At Fn mathlogic, I built an XGBoost credit risk model using a bureau dataset of 90,000+ users. I engineered 234 features and screened them to 33 predictive drivers, achieving 0.68 ROC-AUC on a held-out, out-of-time set.
I also built a 19-step waterfall screening for CMS eligibility, reducing the base from 23.1 lakh live accounts to 15.5 lakh after rule-based exclusions. I tracked counts at every step and built a GST turnover crosstab of the eligible base.
For fraud and account analytics, I built a Power BI dashboard to track fraud counts by city, turnover band, and other KPIs. I defined fraud-incidence measures across geographies and account segments to help identify higher-risk profiles.
Earlier, as a Data Science & Engineering Intern on the Vision Transformer Project, I implemented and benchmarked a ViT for CIFAR-10, reaching 82% accuracy and 95% top-5 accuracy. My projects include SQL and Power BI sales analytics, telecom churn analysis, and a PyTorch plant disease classifier.

