Point-in-time feature serving without actuarial data leakage

FEATSRV: Point-in-Time Correct Feature Serving for Insurance Without Data Leakage

Data leakage is the predominant cause of failure in insurance ML models. FEATSRV guarantees that training and serving use only information known at prediction time — never future data. The system combines Feast, Redis, PySpark, and FastAPI for offline batch PIT joins and online Redis lookups, secured via Cloudflare Tunnel, with a companion CLI and browser dashboard and a synthetic Nimbus insurance dataset that exhibits realistic SCD Type 2 temporal complexity.