# 📋 **README.md**
```
# Dataset Link - https://www.kaggle.com/datasets/ealaxi/paysim1

# Real-Time Fraud Detection System ( FinSecure AI )

An end-to-end machine learning system for detecting fraudulent financial transactions in real-time with 99.98% accuracy.

## Project Overview

Built a production-ready fraud detection system that analyzes transaction patterns using machine learning to identify fraudulent activities with high precision and recall. The system includes a REST API and web interface for real-time fraud detection with AI-powered explanations.

### Key Features
- Real-time fraud detection with less than 100ms latency
- 99.98% AUC-ROC accuracy
- 80% precision at 99% recall
- GPT-powered natural language explanations
- Production-ready REST API
- Interactive web interface

## Performance Metrics

| Metric | Score |
|--------|-------|
| AUC-ROC | 99.98% |
| Precision | 80.34% |
| Recall | 99.20% |
| Training Data | 6.3M transactions |

## Tech Stack

- ML Framework: XGBoost
- API: FastAPI
- AI Explanations: OpenAI GPT-3.5
- Data Processing: Pandas, NumPy
- Deployment: Uvicorn

## Quick Start

### Installation

1. Clone the repository

2. Install dependencies
```
pip install -r requirements.txt
```

3. Set up environment variables
```
echo "OPENAI_API_KEY=your-openai-api-key-here" > .env
```

4. Run the API server
```
uvicorn api:app --reload
```

5. Access the application
- Web Interface: http://127.0.0.1:8000/
- API Documentation: http://127.0.0.1:8000/docs

## How It Works

### Data Pipeline
- Processed 6.3M transaction records
- Time-based train/validation/test split (70/15/15)
- Engineered 15 behavioral features

### Feature Engineering
Key features include:
- Balance Changes: oldbalanceOrg, newbalanceOrig (account draining detection)
- Transaction Patterns: amount, log_amount, recency_hours
- Behavioral Flags: is_dest_new, txn_count_24h
- Transaction Type: Binary flags for TRANSFER, CASH_OUT, PAYMENT, etc.

### Model Training
- Algorithm: XGBoost Classifier
- Handled severe class imbalance (0.13% fraud rate)
- Optimized threshold tuning (0.5 to 0.8)

### Deployment
- REST API with FastAPI
- Real-time scoring endpoint
- GPT-powered explanations for fraud analysts

## API Usage

### Predict Fraud

Endpoint: POST /predict

## Key Insights

1. Account Draining Pattern: Fraudsters typically drain accounts (high balance to zero balance)
2. Transaction Type: TRANSFER and CASH_OUT are highest risk
3. New Destinations: Transfers to unknown recipients are suspicious
4. Rapid Activity: Multiple transactions in short time windows indicate fraud

## Future Improvements

- Add model monitoring and drift detection
- Implement A/B testing framework
- Add SHAP values for detailed explanations
- Integrate with real-time transaction streams
- Add user feedback loop for continuous learning

## Author
Jayesh Vishwakarma
LINKEDIN - www.linkedin.com/in/cmd-jayesh