Shyam Khamkar

• 1. Marginal Customer Segmentation for Risk Mitigation
  
  Technologies: Databricks SQL, Python, PySpark, Decision Trees, WOE/IV, Optimal Binning
  
  Designed and implemented a data-driven customer segmentation framework to identify and flag high-risk (marginal) segments in ongoing sourcing for unsecured lending products (e.g., Personal Loans)
  Evaluated current portfolio KPIs against monthly benchmarks to define "bad" behavior aligned with business policies
  Created a sampling framework using dependent KPI variables and over 17,000 TTD variables from OFFUS and ONUS feature marts
  Applied advanced feature reduction techniques (IV, WOE, Optimal Binning, Variable Clustering) to select top 300–400 variables for model input
  Built and validated a decision tree model on out-of-time datasets to isolate segments with 1.5x higher bad rates than monthly benchmarks
  Proposed high-risk segments for sourcing cutbacks, resulting in 2–5% monthly swap-out savings and improved portfolio health
  Contributed to a ₹15 Cr cost savings in Personal Loan segment through proactive identification of marginal customer cohorts

2. Acquisition Risk Scoring Model Development

Technologies: Python, XGBoost, Boruta, SHAP, SQL, ModelOps

Developed a machine learning–based risk scoring model for customer acquisition using XGBoost, aimed at improving early risk assessment during sourcing
Performed feature selection using Boruta algorithm to identify key predictive variables from a high-dimensional dataset
Conducted hyperparameter tuning to optimize model performance and prevent overfitting
Translated model output probabilities into a risk scorecard, ensuring interpretability and ease of integration with business rules
Evaluated model using statistical metrics including KS, AUC, Gini, and monitored stability using Population Stability Index (PSI)
Employed SHAP values to explain feature contributions and support model transparency for stakeholders and compliance teams
Successfully integrated the model into existing acquisition workflows and decision systems, improving risk discrimination at point-of-sale