🛡️ Anti-Fraud Detection API

Hybrid real-time fraud detection microservice Redis Velocity Check + LightGBM + A/B Testing Router — production-ready architecture

---

🚀 Live Demo

Endpoint	Link
🏠 API Root	fraud-detection-api-1-hf78.onrender.com
📖 Swagger UI	fraud-detection-api-1-hf78.onrender.com/docs
❤️ Health Check	fraud-detection-api-1-hf78.onrender.com/health
💳 Test Cards	fraud-detection-api-1-hf78.onrender.com/cards

---

📊 Architecture — Hybrid Fraud Guardian

Incoming Transaction (card_id + amount only)
        │
        ▼
┌─────────────────────────────────────────┐
│   Layer 1: Redis Velocity Check         │
│   Checks transaction frequency          │
│   Protection against brute force        │
└──────────────────┬──────────────────────┘
                   │
                   ▼
┌─────────────────────────────────────────┐
│   Redis Feature Store                   │
│   Client sends only card_id + amount    │
│   API fetches full historical profile   │
│   (V1-V28 features) from Redis          │
└──────────────────┬──────────────────────┘
                   │
                   ▼
┌─────────────────────────────────────────┐
│   A/B Testing Router                    │
│   80% → Model A (Champion)              │
│   20% → Model B (Challenger)            │
│   Safe testing of new hypotheses        │
└──────────────────┬──────────────────────┘
                   │
                   ▼
┌─────────────────────────────────────────┐
│   LightGBM Inference                    │
│   Class Imbalance handling              │
│   Fraud probability → APPROVE / BLOCK   │
└─────────────────────────────────────────┘

---

🛠 Tech Stack

Component	Technology
Machine Learning	LightGBM (Class Imbalance handling)
Feature Store	Redis (Upstash — in-memory profiles)
Backend	FastAPI, Pydantic
A/B Testing	Custom 80/20 Champion/Challenger Router
Containerization	Docker, Docker Compose
Deployment	Render
Dataset	MLG-ULB Credit Card Fraud (Kaggle)

---

🔑 Key Features

1. Redis Feature Store

Client sends only card_id and amount — the API fetches the full historical profile (V1-V28 features) from Redis automatically. This mirrors production anti-fraud systems at Visa and Mastercard:

# Client sends minimal data
{"card_id": "card_0001", "amount": 500.00}

# API enriches with Redis profile internally
features = cache.get(f"profile:{card_id}")  # V1-V28 from Redis
features["Amount"] = amount                  # Add current transaction

2. A/B Testing Router (Champion/Challenger)

Built-in traffic splitter tests new model versions safely in production:

if random.random() < 0.20:
    active_model = ml_models["model_b"]  # Challenger — 20%
else:
    active_model = ml_models["model_a"]  # Champion — 80%

No downtime, no risk — new models proven before full rollout.

3. Class Imbalance Handling

Credit card fraud datasets are heavily imbalanced (0.17% fraud). LightGBM trained with scale_pos_weight to properly detect rare fraud events.

---

⚡ API Reference

GET /

{
  "service": "Anti-Fraud Detection API",
  "architecture": "Hybrid: Redis Feature Store + LightGBM",
  "features": {
    "ab_testing": "80/20 Champion/Challenger split",
    "feature_store": "Redis in-memory profiles (Upstash)",
    "model": "LightGBM with Class Imbalance handling"
  }
}

GET /health

{
  "status": "healthy",
  "redis": "connected",
  "models_loaded": ["model_a", "model_b"],
  "ab_split": "80% Model A / 20% Model B"
}

GET /cards

{
  "available_cards": ["card_0001", "card_0002", "..."],
  "total": 101,
  "test_cards": {
    "normal": "card_0001",
    "fraud": "fraud_card_001"
  }
}

POST /scan

Normal transaction:

// Request
{
  "card_id": "card_0001",
  "amount": 500.00
}

// Response
{
  "card_id": "card_0001",
  "amount": 500.0,
  "decision": "APPROVE",
  "risk_level": "LOW",
  "risk_probability": 0.0821,
  "model_used": "Model A (Champion)",
  "feature_store": "Redis (Upstash)"
}

Fraud transaction:

// Request
{
  "card_id": "fraud_card_001",
  "amount": 150000.00
}

// Response
{
  "card_id": "fraud_card_001",
  "amount": 150000.0,
  "decision": "BLOCK",
  "risk_level": "HIGH",
  "risk_probability": 0.8934,
  "model_used": "Model A (Champion)",
  "feature_store": "Redis (Upstash)"
}

---

🚀 Quick Start (Local)

1. Clone the repository

git clone https://github.com/RaNurbekov/fraud-detection-api.git
cd fraud-detection-api

2. Start with Docker Compose

docker-compose up --build

This starts both Redis and FastAPI containers automatically.

3. Test the API

# Normal transaction
curl -X POST "http://localhost:8000/scan" \
  -H "Content-Type: application/json" \
  -d '{"card_id": "card_0001", "amount": 500}'

# Fraud transaction
curl -X POST "http://localhost:8000/scan" \
  -H "Content-Type: application/json" \
  -d '{"card_id": "fraud_card_001", "amount": 150000}'

---

📁 Project Structure

fraud-detection-api/
├── api.py                  # FastAPI: /scan, /health, /cards
├── src/
│   └── database.py         # SQLite audit logging
├── models/
│   ├── lgbm_fraud.pkl      # Champion model (Model A)
│   └── lgbm_fraud_b.pkl    # Challenger model (Model B)
├── notebooks/              # Training & EDA
├── docker-compose.yml      # Redis + FastAPI orchestration
├── Dockerfile
├── requirements.txt
└── README.md

---

🔮 Roadmap

Feature	Description
Streamlit Dashboard	Visual A/B testing stats + fraud monitoring
MLflow Integration	Model versioning and experiment tracking
Evidently AI	Data drift monitoring in production
PostgreSQL	Replace SQLite with production-grade DB
Kafka Integration	Connect with kafka-fraud-streaming pipeline

---

🔗 Related Projects

Part of a Fintech ML ecosystem:

• fraud-gnn — Graph Neural Networks for collaborative fraud detection
• credit-risk-api — Credit scoring with MLflow + SHAP + Evidently AI
• kafka-fraud-streaming — Real-time Kafka streaming pipeline

💡 Production vision: Kafka streams transactions → this API scores them in real-time → GNN catches collaborative fraud rings → credit-risk-api assesses customer risk. That's the complete fintech ML stack.

---

📫 Author

Rashid Nurbekov — ML Engineer | Fintech & Generative AI | Almaty, Kazakhstan 🇰🇿