At HealthKart, I worked on The Risk Ledger, an explainable risk-scoring engine for fraud and RTO detection in Indian D2C e-commerce. I engineered an end-to-end pipeline with synthetic data generation, feature engineering, XGBoost training with SHAP explainability, FastAPI inference, and a React analytics dashboard.
I calibrated training data to Indian D2C baselines and achieved 82% precision in the Auto-Block tier. I also designed a threshold simulator that models business impact using measured test-data outcomes, and added automated fraud-pattern detection and a pincode-level risk view.
On my World Happiness Score Prediction project, I built a supervised learning pipeline using multi-year report data and evaluated linear regression predictions. I also implemented numerically stable softmax and cross-entropy loss from scratch while studying neural-network optimization with PyTorch.

