Project Velocity

Title: "Project Velocity": A Holistic AI Underwriting Engine for Car Finance

• Description: I architected and led the development of "Project Velocity," an advanced underwriting platform designed to disrupt the traditional car finance market. The system goes beyond standard CIBIL scores and income statements by integrating a wide array of alternative and proxy data to build a more nuanced and accurate risk profile for each applicant. The core of the platform is a dual-engine system: an Intelligent Document Processing (IDP) module automates the extraction of data from financial documents, while a Holistic Risk Scoring (HRS) module uses a proprietary model that blends traditional credit data with proxy variables, such as geo-demographic data based on the applicant's residence, to better predict creditworthiness, especially for new-to-credit customers.

• Key Features & Innovations:

  • Automated Document Analysis: Used custom models (e.g., LayoutLM, Donut) integrated into a BPM workflow to instantly digitize and verify bank statements, salary slips, and KYC documents, reducing manual effort by over 90%.

  • Alternative Data Engine: Engineered a feature pipeline that ingests and processes non-traditional data. This includes residence locality characteristics (e.g., local economic indicators, property value trends, urban development tier) to serve as a proxy for financial stability and lifestyle.

  • Unified Risk Modeling: Developed a gradient-boosted model (XGBoost) that combines structured data (CIBIL score, transaction history) with the engineered proxy data features, resulting in a significantly more predictive and robust credit risk score.

  • Explainable AI (XAI) for Compliance: Integrated SHAP (SHapley Additive exPlanations) to ensure the model's decisions were transparent and fair. This was critical for demonstrating to regulators that proxy data was used to enhance inclusion, not to discriminate.

• Technologies Used:

  • Python, PyTorch/TensorFlow, Hugging Face Transformers

  • OpenCV, Tesseract for IDP

  • Scikit-learn, XGBoost, GeoPandas for modeling and geospatial feature engineering

  • SHAP, LIME for model explainability

  • FastAPI, Docker, and integration with a Camunda BPM engine

• Outcome: The platform enabled the finance provider to safely approve loans for a 15% larger segment of "thin-file" or new-to-credit applicants, expanding their market share. It also led to a 10% reduction in early-stage defaults within the first year by more accurately flagging high-risk profiles that traditional scoring would have missed.

• GitHub Link that demonstrates the feature engineering process.