I will build fraud detection and credit risk models

Most fraud models are black boxes. Compliance teams, risk officers, and auditors need to know why a transaction was flagged not just that it was.

I build fraud detection systems with SHAP explainability so every model decision has a feature-level audit trail. Defensible for KYC/AML reporting, regulatory review, and credit risk committees.

My work uses XGBoost and LightGBM on imbalanced transaction data. I engineer balance-drain signatures, velocity features, and account deviation indicators. PR-AUC is my primary metric accuracy is meaningless at 1.8% fraud rate.

Built on PaySim: 200K transactions, 31 engineered features, 0.674 PR-AUC baseline actively improving. Real FinTech domain engineering, not toy notebooks.

Services: binary fraud classifier, feature engineering pipeline, SHAP integration, threshold optimisation, model evaluation report (PR curve, ROC, confusion matrix), FastAPI deployment blueprint.

Deliverables: Python notebook or script, trained model artefact, SHAP plots, evaluation report, and audit memo.

Message me with your dataset format, transaction volume, fraud rate, and whether FastAPI deployment is needed.