Cost-sensitive credit card fraud detection on 1.3M transactions. Picks the decision threshold that minimizes expected dollar loss, avoiding $1.25M vs a naive baseline (XGBoost, PR-AUC 0.91). Adds a quant-risk layer (VaR/Expected Shortfall, walk-forward backtesting, PSI drift) and an agentic RAG investigation copilot. Interactive Streamlit app
-
Updated
Jul 31, 2026 - Python