Transaction Fraud Detector

A fraud model needs more than a good score: it needs gated promotion, safe rollout and retraining when data drifts. This project covers that whole lifecycle.

This is an end-to-end MLOps project built in 10 stages. It validates and preprocesses transaction data, trains XGBoost and logistic regression models tracked in MLflow, and only promotes a model if it passes threshold checks and beats the current champion. The promoted model is served with FastAPI and Pydantic validation.

Around the model sit shadow mode, a canary simulation that routes 5% of traffic to the challenger, A/B analysis, Evidently drift reports, a rolling performance monitor, and a retraining pipeline triggered by drift or degradation. The whole thing is Dockerized with CI on GitHub Actions.