Agentic AI Fraud Detection System

An end-to-end AI fraud detection platform combining machine learning, explainable AI, LLM-powered investigation, and a real-time monitoring dashboard.

This project demonstrates how modern financial institutions detect and investigate suspicious transactions using multi-agent AI systems.

---

Key Highlights

• Machine Learning Fraud Detection Model (XGBoost)
• Explainable AI using SHAP
• LLM-powered Investigation Agent
• Multi-Agent Architecture
• FastAPI Fraud Detection API
• Real-time Monitoring Dashboard (Streamlit)

The system automatically:

1. Detects suspicious transactions
2. Explains why they are suspicious
3. Generates investigation summaries
4. Displays alerts in a monitoring dashboard

---

System Architecture

Transaction
     │
     ▼
Detection Agent (XGBoost ML Model)
     │
     ▼
Explainability Engine (SHAP)
     │
     ▼
Investigation Agent (LLM)
     │
     ▼
Decision Agent (Fraud Policy Engine)
     │
     ▼
Fraud Monitoring Dashboard

---

Technology Stack

Layer	Technology
ML Model	XGBoost
Explainability	SHAP
API	FastAPI
Dashboard	Streamlit
LLM Agent	OpenAI
Data Processing	Pandas / NumPy
Deployment	Uvicorn

---

Multi-Agent Design

This project implements 3 cooperating AI agents.

1. Detection Agent

• Predicts fraud probability using XGBoost
• Uses behavioral features such as velocity and spending patterns

2. Investigation Agent

• Uses LLM reasoning to explain suspicious transactions
• Converts SHAP signals into analyst-friendly explanations

Example:

Transaction flagged because the transaction amount is significantly higher than the user's typical spending behavior and multiple transactions occurred within a short time window.

3. Decision Agent

Applies fraud policy rules:

score > 0.7 → BLOCK
0.4–0.7 → INVESTIGATE
< 0.4 → APPROVE

---

Monitoring Dashboard

The Streamlit dashboard allows fraud analysts to:

• Input transaction details
• View fraud risk scores
• See automated investigation summaries
• Visualize risk features

Example output:

Fraud Probability: 0.82
Decision: BLOCK
Reason: Unusual transaction amount and abnormal transaction velocity

---

Fraud Features Used

The model uses behavioral features such as:

TransactionAmt
txn_count_1h
txn_count_24h
txn_count_7d
avg_amt_1h
avg_amt_24h
avg_amt_7d
max_amt_24h
amount_zscore_24h
velocity_risk
is_night_txn

These capture transaction velocity, spending deviation, and behavioral anomalies .

---

Example API Request

POST /predict_fraud

Request:

{
  "TransactionAmt": 500,
  "txn_count_1h": 5,
  "txn_count_24h": 20,
  "txn_count_7d": 40,
  "avg_amt_1h": 30,
  "avg_amt_24h": 35,
  "avg_amt_7d": 28,
  "max_amt_24h": 70,
  "amount_zscore_24h": 4.5,
  "velocity_risk": 0.85,
  "is_night_txn": 1
}

Response:

{
 "fraud_probability": 0.83,
 "decision": "BLOCK",
 "investigation_summary": "The transaction amount is significantly higher than the user's normal spending pattern and multiple transactions occurred within a short time window."
}

OpenAI Multi-Agent Response

---

How to Run the Project

1. Install dependencies

pip install -r requirements.txt

---

2. Start the Fraud Detection API

uvicorn api.app:app --reload

---

3. Launch the Dashboard

streamlit run dashboard/dashboard.py

Open:

http://localhost:8501

---

Future Improvements

• Real-time streaming fraud detection
• Fraud graph network analysis
• Reinforcement learning fraud policies
• Automated case management system

---

Author

Kirti Sinha

Data AI Professional

Passionate about building AI systems that solve real business problems in finance and banking.