Mar 2025 – Jun 2025 · 3 mos · Delhi, India

• Developed ensemble machine learning models using XGBoost, Random Forest, and Neural Networks to detect payment fraud, achieving 89% precision and reducing false positive rates by 40% (from 12% to 7.2%) across 540K+ transaction dataset
• Built real-time risk scoring algorithms processing 2,000+ transactions per second with sub-100ms latency using Apache Kafka and Flink, improving transaction approval rates from 96.8% to 98.7% and enhancing user experience
• Engineered 47 quantitative features including transaction velocity, behavioral patterns, and statistical deviations using PySpark, implementing SMOTE-ENN techniques to handle class imbalance in fraud detection models
• Designed comprehensive merchant risk assessment framework incorporating 15 variables across financial health, operational risk, and behavioral indicators, enabling proactive identification and monitoring of high-risk merchants (10% of portfolio)
• Implemented SHAP-based model interpretability analysis revealing transaction velocity (22% impact) and amount deviation (18% impact) as key fraud predictors, supporting data-driven risk management decisions for millions of monthly UPI transactions