Data Scientist — Credit risk & underwriting at Khatabook (2026-03 – Present)
- Lifted model approval rate 49% to 74% while holding Bucket-A ECL near 3%, architecting the production scoring engine: a 9-model XGBoost ensemble (8 delinquency predictors + a good-payer model) mapped to A–E buckets via KMeans (k=90) calibrated to Expected Credit Loss, with monthly PSI/CSI and KS checks guarding score stability.
- Unlocked lending to previously ineligible sub-6-month-tenure users with an ML whitelisting model on SMS and bureau signals: A/B-quality share of the top cohort 60% to 80%, approval rate 50% to 67% on the top 40% of users.
- Cut manual SQL investigation across a 4,100-table warehouse with a GraphRAG turning questions like "why did approval rate drop?" into validated Snowflake SQL, retrieving join recipes, metric definitions, RCA runbooks from a Neo4j graph of ~2,300 nodes and ~6,900 edges.
- Built and owns the underwriting-monitoring platform (Streamlit + Snowflake) tracking the full credit funnel from BRE policy and model approval through offer, disbursal and DPD/ECL outcomes, with an Airflow + Parquet pipeline delivering zero-touch twice-daily refreshes and daily Slack risk summaries to the risk team.
Data Scientist — Textract Engine for Fortune 500 clients at EXL Services (2024-10 – 2026-02)
- Reduced contract review from 48 hours to under 5 minutes per document for Clorox — a production RAG pipeline extracting 78+ labels across 100+ vendors at 95%+ accuracy with layout-aware OCR and hierarchical indexing.
- Improved contextual precision 30%, reducing hallucinations via context-driven chunking, 2-tier indexing and LLM re-ranking; automated invoice extraction with AWS Textract into unified annual costing sheets.
Data Science Intern at PRAVARTAK, IIT Madras (2023 – 2024)
- Built a CNN classifying respiratory lung sounds from Mel-spectrogram & MFCC audio features with data augmentation, and used SHAP and LIME to make its predictions interpretable for clinical review, achieving robust multi-class disease classification across imbalanced recording conditions.
- Constructed a CNN for leaf-disease classification (Early Blight, Late Blight, Healthy) with data augmentation, Adam optimizer and cross-entropy loss, achieving over 99% validation accuracy at 0.0064 loss.
