Fraud Triage
Fraud models are easy to train and hard to act on. This system turns a LightGBM fraud probability into a decision (approve, send to an analyst, or block) by comparing expected dollar cost under a fixed daily review budget, then gives analysts a review app with the reasons behind each score.
- Leak-free by test. Time-based splits only, and a test proves no feature can see the future.
- Calibrated scores. Platt scaling on a held-out month so probability × amount means real dollars.
- Runs on free hosting. Scoring is offline; the API serves precomputed results and a test checks it never imports the model.