I Trained Six Models for Fraud Detection, and the Best One Isn't in Production

An AI-based fraud detection system was developed by training six machine learning models, including XGBoost and LightGBM, on a combination of two distinct transaction datasets. The process involved creating a unified schema, handling severe class imbalance with SMOTE, and systematically evaluating each model on metrics like AUC-PR. Despite the baseline LightGBM model achieving the highest performance, a tuned XGBoost model was the one put into production, revealing a common disconnect between offline evaluation and deployment choices. The project also evolved to include model explainability using SHAP and a regulatory review workflow to address the steps following a fraud alert.