Tushar Kant Sao

iOS Credit Policy Improvement using Alternative Data •Identified bias in credit models due to missing SMS/app signals for iOS users, leading to underestimation of creditworthiness. •Replaced missing SMS features with equivalent bank statement features to ensure consistent model inputs. •Redesigned underwriting journey to mandate bank statement collection before final decisioning. •Improved model performance (+3% AUC) and approval rates (+5 %). •Unlocked ~₹30Cr monthly disbursal uplift through improved risk segmentation using updated model scores. Loan Amount Optimization (Credit Limit / EMI Logic) •Derived repayment capacity using Experian data based on customers’ past loan repayment behavior •Estimated maximum affordable EMI from historical payments, adjusted for customer risk and existing obligations ( last 2 months active EMI ) to assign appropriate credit limits •Transitioned to CIBIL-based approach (default bureau pull for all customers) and incorporated ROI/tenure inferencing to handle missing data and improve fill rates •Delivered ~10% increase in average ticket size while maintaining stable bad rate. Feature Engineering •Engineered 200+ features (post initial screening) across bureau data, bank statements, derived transformations, and clickstream signals for underwriting and credit risk models. •Evaluated and shortlisted features using IV, KS, correlation, and SHAP importance •Top 2 new bureau features contributed ~15% feature importance in core bureau underwriting model Improved Underwriting Utilizing Lender Information •Developed a separate underwriting policy for rejected customers using a DTC data (bureau data enriched with lender-level information) based model in collaboration with DS team •Optimized model cutoffs (CMA-ES), driving ~₹12Cr monthly disbursal uplift (applied to 40% of new Navi customers) while maintaining Bad rate •Identified optimal bad rate metric for policy creation using F1-score & unit economics-based profitability analysis.