Bhavya SoniIT Undergraduate · ML & MLOps Engineer

Production-style MLOps system trained on 2.26M Lending Club records. Eliminated 14 post-loan target leakage features (debt collections, recovery history) to establish an honest 0.71 ROC-AUC baseline on application-time signals only (preventing inflated 0.95+ leakage scores). Dual FastAPI endpoints (/predict and /explain) served via Docker with Kubernetes self-healing and Prometheus/Grafana drift monitoring.

Low-latency fraud scoring REST API addressing extreme class imbalance (only 0.16% fraud in 94K transactions). Applied SMOTE synthetic oversampling with XGBoost, achieving 0.98 ROC-AUC with 87% recall and 82% precision (flagged 127 of 149 frauds at only 35 false alarms).