Skip to content
All projects
Case study

CardShield

Real-time fraud detection service that scores every card transaction

github.com/himaallu/CardSheild

01The problem

Real-time fraud detection service that scores every card transaction.

02What I built

  • Cut total fraud cost by 65% vs the best amount rule (recall 0.80) by setting the decision threshold by business cost.
  • Served LightGBM via FastAPI at 13 ms p95 with per-decision SHAP reasons, after comparing 4 models on a time-based split.
  • Tracked models in MLflow and added drift monitoring that caught all 3 injected shifts with no false alarms; Docker, 50 tests and CI.

03Architecture