Sentinel

01The problem

Fraud detection needs both real-time inference and a reason a human can read. Black-box scores alone don't cut it when a declined transaction gets reviewed.

02Approach

• 01Built a max-vote ensemble of Random Forest, XGBoost, and LightGBM for the scoring layer.
• 02Served FastAPI prediction endpoints with in-memory velocity rate limiting.
• 03Applied heuristic overrides to hard-rule obvious cases.
• 04Reported Euclidean anomaly tracing surfaced per-prediction explanations (XAI).

03Outcomes

• Ensemble reported ~0.9995 ROC AUC on a 1.27M-sample test set (PaySim).
• Rate limiting + overrides keep the API honest under bursts.
• Anomaly tracing turns predictions into auditable explanations.