ClaimLens AI

AI-powered insurance fraud detection using ML, computer vision, graph analytics, and LLMs.

What is this?

ClaimLens analyzes insurance claims to detect fraud using multiple AI techniques:

• ML Engine: CatBoost classifier with 145 features
• Computer Vision: Document forgery detection (PAN, Aadhaar, licenses, etc.)
• Graph Analytics: Fraud network detection using Neo4j
• LLM: Natural language explanations powered by Groq

The system can process claims in English and Hinglish (Hindi-English mix).

Quick Start

Requirements:

• Python 3.10+
• Docker (optional, for Neo4j)
• Groq API key (free at console.groq.com)

Setup:

git clone https://github.com/pranaya-mathur/ClaimLens_App.git
cd ClaimLens_App

python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate

pip install -r requirements.txt

cp .env.example .env
# Add your GROQ_API_KEY to .env

Run:

# Start backend
uvicorn api.main:app --reload

# Start frontend (new terminal)
streamlit run frontend/streamlit_app.py

# Optional: Neo4j for graph features
docker-compose up neo4j -d

Visit http://localhost:8501 for the UI or http://localhost:8000/docs for API docs.

How it Works

The system processes claims through four engines:

1. ML Engine extracts 145 features (text embeddings, behavioral patterns, document flags) and runs CatBoost classification
2. CV Engine verifies document authenticity using ResNet50 + Error Level Analysis
3. Graph Engine checks for fraud networks (shared documents, suspicious connections)
4. LLM Engine aggregates results and generates explanations

All results are combined using semantic aggregation to produce a final verdict.

API Examples

Score a claim:

POST /api/ml/score
{
  "claim_id": "CLM001",
  "narrative": "Car accident on highway",
  "claim_amount": 50000,
  "product": "motor",
  ...
}

Verify a document:

POST /api/documents/verify-pan
(upload PAN card image)

Complete analysis:

POST /api/unified/analyze-complete
(runs all engines)

See /docs folder for detailed API documentation.

Features

• Multi-modal fraud detection
• Document forgery detection with OCR
• Fraud network visualization
• AI-generated explanations (technical & customer-friendly)
• Real-time processing (<2s per claim with caching)
• Hinglish support

Project Structure

ClaimLens_App/
├── api/              # FastAPI backend
│   ├── routes/       # API endpoints
│   └── schemas/      # Pydantic models
├── src/              # Core engines
│   ├── ml_engine/    # CatBoost fraud scoring
│   ├── cv_engine/    # Document verification
│   ├── llm_engine/   # LLM explanations
│   └── fraud_engine/ # Graph analytics
├── frontend/         # Streamlit UI
├── models/           # Trained models
├── docs/             # Documentation
└── tests/            # Test suite

Configuration

Key environment variables (see .env.example):

GROQ_API_KEY=your_key_here
NEO4J_URI=bolt://localhost:7687
NEO4J_PASSWORD=your_password

ENABLE_LLM_EXPLANATIONS=true
ML_THRESHOLD=0.5

Performance

Based on testing with synthetic claims:

• ML inference: ~80ms
• Document analysis: 1-2s
• LLM explanations: 2-3s
• End-to-end (cached): <500ms

Rate limited to 100 requests/min by default.

Known Issues

• Neo4j sometimes needs a restart on first launch
• Large PDFs (>10MB) may timeout
• Hinglish embeddings need mixed English-Hindi text to work well
• Generic document detector accuracy varies by document quality

See Issues for more.

TODO

• Add batch processing endpoint
• Support PDF documents
• Add more languages (Hindi, Tamil, Telugu)
• Improve graph visualizations
• Add model retraining pipeline

See full roadmap in CHANGELOG.md.

Contributing

PRs welcome! Please:

1. Fork the repo
2. Create a feature branch
3. Add tests for new features
4. Submit a PR

See CONTRIBUTING.md for guidelines.

License

MIT - see LICENSE

Contact

Pranay Mathur

• GitHub: @pranaya-mathur
• Email: pranaya.mathur@yahoo.com

Built with help from Groq, LangChain, Neo4j, and CatBoost.