🛡️ Fraud Detection Agent - AI-Powered Transaction Risk Assessment

A sophisticated LangGraph-based fraud detection system that leverages parallel AI analysis to detect fraudulent transactions in real-time. Built with Python, FastAPI, and OpenAI's language models, this agent performs multi-dimensional risk assessment through five specialized analyzers running in parallel.

🎯 Overview

What is this Agent?

The Fraud Detection Agent is an intelligent system that analyzes financial transactions for potential fraud using multiple AI-powered analyzers working in parallel. It provides real-time risk assessment with three possible outcomes: APPROVE, REVIEW, or DECLINE.

Key Capabilities

Parallel Analysis: 5 specialized fraud detectors run simultaneously
LLM-Powered: Uses OpenAI's GPT models for sophisticated pattern recognition
Flexible Input: Accepts multiple transaction data formats
Production-Ready: Deployed on Google Cloud Run with monitoring
Configurable: Adjustable risk thresholds and analyzer weights
Fault-Tolerant: Built-in retry logic and timeout protection

Technology Stack

Orchestration: LangGraph (state-based parallel workflow)
AI/ML: OpenAI GPT-4o-mini (configurable)
Runtime: FastAPI + Uvicorn
Language: Python 3.13
Deployment: Google Cloud Run
Monitoring: Handit AI

🏗️ Architecture

High-Level System Architecture

┌─────────────────────────────────────────────────────────────┐
│                     FastAPI Server (Port 8001)              │
│                    with Handit AI Tracing                   │
└────────────────┬────────────────────────────────────────────┘
                 │
                 ▼
        ┌────────────────────┐
        │   LangGraphAgent   │
        │  (Main Orchestrator)│
        └────────────┬───────┘
                     │
        ┌────────────▼────────────┐
        │  RiskManagerGraph       │
        │  (StateGraph Based)     │
        └────────────┬────────────┘
                     │
        ┌────────────▼──────────────────────────────────────┐
        │        ORCHESTRATOR NODE (START)                  │
        │  Normalizes and enriches transaction data          │
        └────────────┬──────────────────────────────────────┘
                     │
        ┌────────────▼──────────────────────────────────────┐
        │         PARALLEL ANALYZER EXECUTION               │
        │  ┌──────────────┐ ┌──────────────┐               │
        │  │  Pattern     │ │ Behavioral   │               │
        │  │  Detector    │ │ Analyzer     │               │
        │  └──────────────┘ └──────────────┘               │
        │  ┌──────────────┐ ┌──────────────┐ ┌──────────┐ │
        │  │  Velocity    │ │ Merchant     │ │Geographic││
        │  │  Checker     │ │ Risk Analyzer│ │ Analyzer ││
        │  └──────────────┘ └──────────────┘ └──────────┘ │
        └────────────┬──────────────────────────────────────┘
                     │
        ┌────────────▼──────────────────────┐
        │  DECISION AGGREGATOR NODE         │
        │  Combines all analyzer results    │
        └────────────┬──────────────────────┘
                     │
                     ▼
        ┌────────────────────────────────┐
        │  Final JSON Decision Output    │
        │  {final_decision, reason, ...} │
        └────────────────────────────────┘

Project Structure

risk_manager/
├── main.py                          # FastAPI application entry point
├── src/
│   ├── agent.py                     # LangGraphAgent orchestrator
│   ├── base.py                      # Base node classes
│   ├── config.py                    # Configuration and graph topology
│   ├── graph/
│   │   ├── main.py                  # RiskManagerGraph implementation
│   │   └── nodes/
│   │       └── nodes.py             # Node function definitions
│   ├── nodes/
│   │   └── llm/                     # LLM-based analyzer nodes
│   │       ├── pattern_detector/    # Fraud pattern detection
│   │       ├── behavioral_analizer/ # Behavioral anomaly detection
│   │       ├── velocity_checker/    # Velocity abuse detection
│   │       ├── merchant_risk_analizer/ # Merchant risk assessment
│   │       ├── geographic_analizer/ # Geographic fraud detection
│   │       └── decision_aggregator/ # Final decision maker
│   ├── state/
│   │   └── state.py                 # AgentState definition
│   └── utils/
│       └── openai_client.py         # OpenAI client wrapper
├── use_cases/                       # Test cases and examples
├── requirements.txt                 # Python dependencies
├── .env.example                     # Environment variables template
└── cloudbuild.yaml                  # Google Cloud Build config

✨ Features

Core Capabilities

1. Parallel Processing Architecture

5 analyzers run simultaneously, not sequentially
Reduces processing time from ~5x to ~1x
Automatic state merging with LangGraph

2. Multi-Format Input Support

Accepts various transaction data formats:

Simple format (basic fields)
Complex banking format (comprehensive data)
Legacy format (backward compatibility)
Hybrid format (mixed structure)

3. Intelligent Risk Assessment

Weighted scoring from multiple analyzers
Configurable risk thresholds
Evidence-based decision making

4. Production-Ready Infrastructure

Async/await throughout
Retry with exponential backoff
Timeout protection (120s max)
Health checks and monitoring

The Five Fraud Analyzers

Analyzer	Purpose	Weight	Key Detection Areas
Pattern Detector	Identifies known fraud signatures	25%	Synthetic identity, account takeover, card testing, money laundering
Behavioral Analyzer	Detects user behavior anomalies	20%	Spending patterns, temporal deviations, merchant preferences
Velocity Checker	Catches rapid-fire attacks	25%	Transaction frequency, amount velocity, failure rates
Merchant Risk Analyzer	Assesses merchant trustworthiness	15%	Merchant category, reputation, fraud history
Geographic Analyzer	Detects location-based fraud	15%	Impossible travel, VPN detection, high-risk regions

🚀 Installation

Prerequisites

Python 3.13+
OpenAI API key
(Optional) Handit AI API key for monitoring
(Optional) Google Cloud account for deployment

Quick Start

Clone the repository

git clone <repository-url>
cd risk_manager

Create virtual environment

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

Install dependencies

pip install -r requirements.txt

Set up environment variables

cp .env.example .env
# Edit .env and add your OPENAI_API_KEY

Run the application

python main.py

The server will start on http://localhost:8001

⚙️ Configuration

Environment Variables

Create a .env file with the following variables:

# Required
OPENAI_API_KEY=sk-...           # Your OpenAI API key

# Optional
HANDIT_API_KEY=...               # Handit AI monitoring (optional)
MODEL_NAME=gpt-4o-mini           # LLM model to use (default: gpt-4o-mini)
MODEL_PROVIDER=openai            # LLM provider (default: mock for testing)
ENVIRONMENT=development          # Environment (development/production)
HOST=0.0.0.0                     # Server host
PORT=8001                        # Server port
LOG_LEVEL=info                   # Logging level (debug/info/warning/error)

Risk Thresholds

Modify risk thresholds in src/config.py:

self.risk_thresholds = {
    "decline": 70,    # Risk score >= 70: DECLINE
    "review": 40,     # Risk score >= 40: REVIEW
    "approve": 0      # Risk score < 40: APPROVE
}

Analyzer Weights

Adjust the importance of each analyzer in src/config.py:

self.analyzer_weights = {
    "pattern_detector": 0.25,       # 25% weight
    "behavioral_analizer": 0.20,    # 20% weight
    "velocity_checker": 0.25,        # 25% weight
    "merchant_risk_analizer": 0.15, # 15% weight
    "geographic_analizer": 0.15      # 15% weight
}

📡 Usage

Basic API Call

curl -X POST http://localhost:8001/process \
  -H "Content-Type: application/json" \
  -d '{
    "user_id": "john_doe_123",
    "amount": 250.50,
    "merchant_name": "Amazon",
    "user_age_days": 365,
    "total_transactions": 150
  }'

Python Example

import requests

# Transaction data
transaction = {
    "user_id": "customer_123",
    "amount": 500.00,
    "merchant_name": "Electronics Store",
    "user_age_days": 180,
    "total_transactions": 50,
    "location": "New York, US",
    "time": "14:30",
    "velocity_counters": {
        "transactions_last_hour": 2,
        "declined_transactions_last_24h": 0
    }
}

# Send request
response = requests.post(
    "http://localhost:8001/process",
    json=transaction
)

# Parse response
result = response.json()
decision = result["result"]["decision"]["final_decision"]
reason = result["result"]["decision"]["reason"]

print(f"Decision: {decision}")
print(f"Reason: {reason}")

Response Format

{
  "result": {
    "pattern_detector": "No suspicious patterns detected...",
    "behavioral_analizer": "Transaction aligns with user history...",
    "velocity_checker": "Normal transaction velocity...",
    "merchant_risk_analizer": "Trusted merchant...",
    "geographic_analizer": "Location consistent...",
    "decision": {
      "final_decision": "APPROVE",
      "conclusion": "Low-risk transaction from established user",
      "recommendations": ["Process transaction normally"],
      "reason": "All analyzers indicate low fraud risk..."
    }
  },
  "success": true,
  "metadata": {
    "agent": "risk_manager",
    "framework": "langgraph",
    "processing_time_ms": 3250.45
  }
}

📚 API Documentation

Endpoints

POST /process

Process a transaction through fraud detection

Request Body (Multiple formats supported):

Simple Format:

{
  "user_id": "string",
  "amount": 0.0,
  "merchant_name": "string",
  "user_age_days": 0,
  "total_transactions": 0
}

Complex Banking Format:

{
  "transaction": {
    "transaction_id": "string",
    "amount": 0.0,
    "currency": "USD"
  },
  "merchant": {
    "merchant_name": "string",
    "merchant_category_code": "string"
  },
  "customer": {
    "user_id": "string",
    "user_age_days": 0
  },
  "velocity_counters": {
    "transactions_last_hour": 0,
    "declined_transactions_last_24h": 0
  }
}

Response: ProcessResponse with analyzer results and decision

GET /health

Health check endpoint

Response:

{
  "status": "healthy",
  "agent": "risk_manager",
  "framework": "langgraph"
}

GET /graph/info

Get graph structure information

Response: Graph topology and node information

🔍 Fraud Detection Methodology

Detection Strategy

The system employs a multi-layered defense strategy:

Pattern Recognition: Identifies known fraud signatures
Behavioral Analysis: Detects anomalies from established patterns
Velocity Monitoring: Catches rapid-fire attacks
Merchant Assessment: Evaluates merchant trustworthiness
Geographic Verification: Validates location consistency

Risk Factors Analyzed

Transaction Risk Factors

Amount anomalies (unusually high amounts)
Time-based risks (transactions at 2-5 AM)
New user risks (account age < 7 days)
Authentication failures

Behavioral Risk Factors

Spending pattern deviations
Merchant preference changes
Location inconsistencies
Device/channel changes

Velocity Risk Factors

High transaction frequency
Escalating amounts
Multiple declined attempts
Rapid account changes

Decision Logic

Risk Score Calculation:
├─ Each analyzer returns risk assessment
├─ Scores are weighted by importance
├─ Final score = Weighted average (0-100)
└─ Decision based on thresholds:
    ├─ Score < 40: APPROVE ✅
    ├─ Score 40-69: REVIEW 🔍
    └─ Score >= 70: DECLINE ❌

Example Fraud Scenarios

Scenario 1: Account Takeover

Pattern Detector: "Account takeover signature detected"
Behavioral: "Significant deviation from normal behavior"
Geographic: "Login from different country"
Decision: DECLINE

Scenario 2: Card Testing

Pattern Detector: "Multiple small transactions pattern"
Velocity: "10 transactions in 5 minutes"
Merchant: "Multiple different merchants"
Decision: DECLINE

Scenario 3: Legitimate High-Value Purchase

Pattern Detector: "No fraud patterns"
Behavioral: "Consistent with user profile"
Merchant: "Trusted merchant"
Decision: APPROVE

🔄 LangGraph Implementation

Why LangGraph?

LangGraph provides the perfect framework for our parallel fraud detection:

State Management: Automatic state merging from parallel branches
Graph Structure: Visual workflow representation
Error Handling: Built-in retry and error recovery
Async Support: Native async/await for high performance

Graph Configuration

The graph topology is defined in src/config.py:

graph_config = {
    "nodes": {
        "orchestrator": {...},           # Entry point
        "pattern_detector": {...},        # Parallel analyzer 1
        "behavioral_analizer": {...},     # Parallel analyzer 2
        "velocity_checker": {...},        # Parallel analyzer 3
        "merchant_risk_analizer": {...},  # Parallel analyzer 4
        "geographic_analizer": {...},     # Parallel analyzer 5
        "decision_aggregator": {...},     # Convergence point
        "finalizer": {...}                # Final processing
    },
    "edges": [
        # Fan-out (1 to many)
        {"from": "orchestrator", "to": "pattern_detector"},
        {"from": "orchestrator", "to": "behavioral_analizer"},
        {"from": "orchestrator", "to": "velocity_checker"},
        {"from": "orchestrator", "to": "merchant_risk_analizer"},
        {"from": "orchestrator", "to": "geographic_analizer"},

        # Fan-in (many to 1)
        {"from": "pattern_detector", "to": "decision_aggregator"},
        {"from": "behavioral_analizer", "to": "decision_aggregator"},
        {"from": "velocity_checker", "to": "decision_aggregator"},
        {"from": "merchant_risk_analizer", "to": "decision_aggregator"},
        {"from": "geographic_analizer", "to": "decision_aggregator"},

        # Sequential
        {"from": "decision_aggregator", "to": "finalizer"}
    ]
}

State Management

The AgentState (TypedDict) manages data flow:

class AgentState(TypedDict):
    input: Dict[str, Any]
    transaction_data: Dict[str, Any]
    enriched_transaction: Dict[str, Any]
    analyzer_results: Annotated[Dict, merge_dicts]  # Parallel merge
    risk_scores: Annotated[Dict, merge_dicts]       # Parallel merge
    completed_analyzers: List[str]
    final_decision: str
    results: Annotated[Dict, merge_dicts]           # Parallel merge

Parallel Execution Flow

1. Orchestrator prepares transaction data
2. LangGraph spawns 5 parallel branches
3. Each analyzer runs independently
4. Results automatically merge (Annotated fields)
5. Decision aggregator receives all results
6. Final decision generated

🚢 Deployment

Google Cloud Run Deployment

The application is configured for Google Cloud Run deployment with automatic CI/CD.

Prerequisites

Google Cloud Project
Cloud Build API enabled
Artifact Registry repository created

Deployment Steps

Configure secrets in Google Cloud

echo "your-openai-key" | gcloud secrets create OPENAI_API_KEY --data-file=-

Update cloudbuild.yaml with your project details
Deploy using Cloud Build

gcloud builds submit --config=cloudbuild.yaml

Access the deployed service

gcloud run services describe transaction-validation-agent --region=us-central1

Docker Deployment

Build and run with Docker:

# Build image
docker build -t fraud-detection-agent .

# Run container
docker run -p 8001:8001 \
  -e OPENAI_API_KEY=your-key \
  fraud-detection-agent

Configuration for Production

Recommended Cloud Run settings:

Memory: 4 GiB (for LLM processing)
CPU: 2 vCPUs
Timeout: 600 seconds
Concurrency: 10 requests
Min instances: 0 (scale to zero)
Max instances: 10 (adjust based on load)

⚡ Performance

Benchmarks

Metric	Value	Notes
Average Processing Time	3-4 seconds	For complete analysis
Parallel Speedup	~5x	Compared to sequential
Throughput	15-20 req/min	Per instance
Timeout Rate	< 1%	With 120s timeout
Success Rate	> 99%	With retry logic

Optimization Tips

Reduce Analyzer Count: Disable less critical analyzers for speed
Adjust Timeouts: Lower timeout for faster failure detection
Cache Results: Implement Redis for repeat transactions
Batch Processing: Process multiple transactions together
Model Selection: Use faster models (gpt-3.5-turbo) for lower latency

Resource Requirements

Minimum: 2 GB RAM, 1 vCPU
Recommended: 4 GB RAM, 2 vCPUs
Network: Low bandwidth, ~1 KB per request
Storage: Minimal, logs only

🔧 Troubleshooting

Common Issues and Solutions

1. OpenAI API Errors

Error: OpenAI API key not found Solution: Ensure OPENAI_API_KEY is set in environment

2. Timeout Errors

Error: asyncio.TimeoutError Solution:

Increase timeout in src/graph/main.py
Check OpenAI API status
Reduce parallel analyzer count

3. High Latency

Symptoms: Requests taking > 10 seconds Solutions:

Switch to faster model (gpt-3.5-turbo)
Reduce prompt complexity
Implement caching layer

4. Memory Issues

Error: Out of memory Solution: Increase Cloud Run memory allocation

Debug Mode

Enable debug logging:

LOG_LEVEL=debug python main.py

View detailed traces:

Check Handit AI dashboard (if configured)
Review application logs
Use /graph/info endpoint for graph details

🛠️ Development

Adding New Analyzers

Create analyzer directory

mkdir src/nodes/llm/new_analyzer

Implement processor

# src/nodes/llm/new_analyzer/processor.py
from src.base import BaseLLMNode

class NewAnalyzerLLMNode(BaseLLMNode):
    async def run(self, input_data):
        # Implementation
        pass

Define prompts

# src/nodes/llm/new_analyzer/prompts.py
def get_prompts():
    return {
        "system": "You are an expert...",
        "user_template": "Analyze..."
    }

Register in config

# src/config.py
# Add to graph_config nodes and edges

Add node function

# src/graph/nodes/nodes.py
async def new_analyzer_node(state):
    # Node implementation
    pass

Running Tests

Execute test suite:

# Run all tests
pytest

# Run with coverage
pytest --cov=src

# Run specific test file
pytest tests/test_fraud_detection.py

Running Use Cases

Test with provided use cases:

# Run all use cases
python run_use_cases.py

# Run specific file
python run_use_cases.py --file use_cases/SOPHISTICATED_FRAUD_CASES.json

# Generate report
python run_use_cases.py --report

Code Quality

Maintain code quality:

# Format code
black .

# Lint
flake8 src/

# Type checking
mypy src/

📄 License

🤝 Contributing

Please follow these guidelines:

Fork the repository
Create feature branch (git checkout -b feature/AmazingFeature)
Commit changes (git commit -m 'Add AmazingFeature')
Push to branch (git push origin feature/AmazingFeature)
Open Pull Request

📞 Support

Documentation: Handit.ai Documentation
LangGraph: LangGraph Documentation
Issues: Open an issue in the repository

🙏 Acknowledgments

Built with LangGraph for orchestration
Powered by OpenAI for AI analysis
Monitored by Handit AI for observability
Deployed on Google Cloud Run for scalability

Version: 1.0.0 Last Updated: January 2025 Status: Production Ready

FilesExpand file tree

README.md

Latest commit

History

README.md

File metadata and controls

🛡️ Fraud Detection Agent - AI-Powered Transaction Risk Assessment

📋 Table of Contents

🎯 Overview

What is this Agent?

Key Capabilities

Technology Stack

🏗️ Architecture

High-Level System Architecture

Project Structure

✨ Features

Core Capabilities

1. Parallel Processing Architecture

2. Multi-Format Input Support

3. Intelligent Risk Assessment

4. Production-Ready Infrastructure

The Five Fraud Analyzers

🚀 Installation

Prerequisites

Quick Start

⚙️ Configuration

Environment Variables

Risk Thresholds

Analyzer Weights

📡 Usage

Basic API Call

Python Example

Response Format

📚 API Documentation

Endpoints

POST /process

GET /health

GET /graph/info

🔍 Fraud Detection Methodology

Detection Strategy

Risk Factors Analyzed

Transaction Risk Factors

Behavioral Risk Factors

Velocity Risk Factors

Decision Logic

Example Fraud Scenarios

Scenario 1: Account Takeover

Scenario 2: Card Testing

Scenario 3: Legitimate High-Value Purchase

🔄 LangGraph Implementation

Why LangGraph?

Graph Configuration

State Management

Parallel Execution Flow

🚢 Deployment

Google Cloud Run Deployment

Prerequisites

Deployment Steps

Docker Deployment

Configuration for Production

⚡ Performance

Benchmarks

Optimization Tips

Resource Requirements

🔧 Troubleshooting

Common Issues and Solutions

1. OpenAI API Errors

2. Timeout Errors

3. High Latency

4. Memory Issues

Debug Mode

🛠️ Development

Adding New Analyzers

Running Tests

Running Use Cases

Code Quality

📄 License

🤝 Contributing

📞 Support

🙏 Acknowledgments