Kanpur, Uttar Pradesh, India
Designed and implemented a fraud detection framework using Supervised Contrastive Learning on 500,000+ UPI P2P transactions with 45+ behavioral and transactional features, addressing severe class imbalance (<2% fraud cases). Applied undersampling (1:10 fraud-to-non-fraud ratio) and ensemble techniques — achieved Precision: 37% | Recall: 34% | F1-score: 35%, outperforming XGBoost and class-weighted baselines. Validated model robustness on 450,000 real-world cases, demonstrating scalability for national-level payment systems and enhancing fraud prevention for growing UPI volumes.