I built Fraud Investigation Squad, a live four-agent fraud-triage system combining XGBoost anomaly detection, policy checks, an LLM agent, and signal reconciliation. Using the 6.36M-row PaySim dataset, I achieved 0.9993 ROC-AUC and reduced an over-triggering policy rule from roughly 80% to 0.5% through live percentile-based calibration.
I've also built CNN-based medical imaging and food-freshness systems, including a 14-class chest X-ray classifier that reached 0.89 AUC and a MobileNetV2 freshness classifier with 98.5% validation accuracy. I enjoy taking ML projects end to end, from preprocessing and explainability through FastAPI, React, Streamlit, and production deployment.
