Machine learning prototype for profiling risk in insurance claims using AutoGluon and SHAP explainability.
- Fraud Prediction – Binary classification using trained AutoGluon ensemble models
- Probability Scoring – Configurable fraud probability threshold with risk assessment
- SHAP Explainability – Per-prediction feature contributions showing which inputs drive the result
- Feature Importance – Global feature importance visualization from SHAP training analysis
- AI Summary – Natural language fraud assessment generated via OpenAI
- Interactive Dashboard – Input form with real-time prediction, gauge visualization, and feature impact charts
The system uses an AutoGluon TabularPredictor trained on the 2023 Travelers NESS Statathon dataset (insurance claim records with driver demographics, claim details, and vehicle information). When a user submits claim features through the dashboard, the backend runs the model prediction, computes SHAP contributions using a KernelExplainer with 25-sample background, and generates a natural language summary via GPT-4o-mini. Claims exceeding the 65% fraud probability threshold are flagged as high risk.
| Category | Technologies |
|---|---|
| Backend | Python 3.13, FastAPI, Uvicorn |
| Frontend | TypeScript, Next.js, React, Tailwind CSS |
| AI/ML | AutoGluon, scikit-learn, SHAP, OpenAI |
| Data | pandas, NumPy, Pydantic |
| Package Management | uv (backend), pnpm (frontend) |
| Deployment | Docker, GitHub Actions, Google Artifact Registry |
# Backend
cd backend
cp .env.example .env # Add your OpenAI API key
uv sync
uv run uvicorn risk_profiler.main:app --reload
# Frontend
cd frontend
pnpm install
pnpm devOpen http://localhost:3000 (frontend) and http://localhost:8000/docs (API docs).