FraudGuard: AI-Powered Fraud Detection & Investigation System

FraudGuard is an end-to-end fraud detection and investigation system built on the PaySim dataset, combining Machine Learning, Rule-Based Scoring, and LLM-driven explanations, with an interactive Streamlit dashboard for real-time analysis.

This project demonstrates how modern financial fraud systems move beyond binary flags into explainable, investigator-ready intelligence.

Key Features

Multi-Layer Fraud Scoring

FraudGuard evaluates each transaction using:

• ML Fraud Score (anomaly detection model)
• Rule Engine Score (domain rules: balance mismatch, risky transaction types, etc.)
• Final Fraud Score (weighted ensemble of ML + rules)
• Transactions crossing a risk threshold are automatically flagged.

AI Investigator (LLM)

For high-risk transactions only, FraudGuard generates:

• Risk Level (Low / Medium / High)
• Key Risk Explanation
• Recommended Action

Interactive Streamlit Dashboard

The dashboard allows users to:

• Enter a Transaction ID to view full transaction details

• Instantly see:

  • Fraud scores
  • Balance changes
  • Rule engine reasons
  • AI investigator explanation

Tech Stack

• Python
• Pandas / NumPy
• Scikit-learn
• LLMs via Free APIs - Groq (LLaMA-3.1)
• Streamlit

Dataset

Dataset used: https://www.kaggle.com/datasets/ealaxi/paysim1

Project Architecture

Gen-AI_FraudGuard/
│
├── Codes/                          # Main code directory
│   ├── datapreprocess_model.ipynb   
│   ├── llm_investigations.ipynb            
│   ├── check_for_fraudid.ipynb      
│   └── frontend.py                
│
├── Datasets/                       
│   ├── paysim_processed_with_scores.csv    
│   ├── paysim_processed.csv 
│   └── paysim_sample.csv 
│
└── README.md

Fraud Scoring Logic

Total Fraud Score =
0.6 × ML Fraud Score +
0.2 × Rule Score +
0.2 × LLM Score

Flagging threshold:

Total Fraud Score ≥ 0.7 → Flagged Fraud

Results

This project reflects real-world fraud systems used in fintech and banking:

• Combines statistical ML with deterministic rules
• Adds explainability, not just predictions
• Optimizes LLM usage for cost and relevance
• Focuses on analyst workflow, not just model accuracy