AI-Powered Experience Continuity Engine

Predictive degradation detection module — monitors server telemetry (CPU, memory, HTTP 500 rate, network latency) and flags early-stage degradation before a full outage, triggering a reroute signal for the traffic router.

Model Performance

Model: IsolationForest (unsupervised anomaly detection), random_state=42, n_jobs=1 Threshold: -0.500710 (58th percentile of degrading-segment scores — the most aggressive cutoff that avoids false positives on normal traffic)

Confusion Matrix

	Predicted Normal	Predicted Degrading
Actual Normal (3,600)	3,514 (TN)	86 (FP)
Actual Degrading (1,800)	756 (FN)	1,044 (TP)

Metrics

Metric	Value	What it means
Precision (degrading)	92.4%	When we call it "degrading," we're right 92% of the time — few false alarms
Recall (degrading)	58.0%	We catch 58% of all degrading readings — biased toward precision over recall to avoid rerouting healthy traffic
F1	0.713	Balanced measure of precision/recall
Accuracy	84.4%	Overall correctness (less meaningful here — classes are imbalanced 2:1 normal:degrading)
Detection point	~42% through ramp	Reroute fires with over half the degradation window still remaining before critical state
Inference latency	Sub-millisecond	No I/O or DataFrame overhead — model + scaler loaded once at startup

Pipeline

simulator.py  →  historical_logs.csv  →  train.py  →  model.joblib  →  app.py (FastAPI /predict)  →  demo_client.py

1. simulator.py — generates synthetic telemetry: a healthy baseline segment followed by a smooth, gradual degradation ramp (not a step change), so early detection is possible.
2. train.py — trains IsolationForest on 4 raw features (CPU, memory, error rate, latency), tunes the threshold against the labeled degrading segment, saves model.joblib.
3. app.py — FastAPI service. Loads the model once at startup. POST /predict returns {"status": "healthy" | "degrading", "health_score": <float 0-1>, "action": "none" | "reroute"}.
4. demo_client.py — streams simulated live telemetry to the API once per second, prints color-coded status, and fires a one-time reroute banner when action == "reroute".

Running it

python3 simulator.py      # generates historical_logs.csv
python3 train.py          # trains model, saves model.joblib
python3 app.py             # starts API on :8000

# in a second terminal
python3 demo_client.py