RiskRadar AI

AI-Powered Fraud Transaction Investigation Assistant

IIT Roorkee AIOps Capstone Project — Theme 13

---

Project Overview

RiskRadar AI is an intelligent agentic system that assists fraud analysts in investigating suspicious bank transactions in real time. Built with LangGraph and LangChain, it combines multi-agent reasoning, policy-grounded RAG retrieval, and explainable recommendations inside an interactive Streamlit interface.

The system aims to reduce investigation time while improving consistency, auditability, and decision quality for fraud alerts.

---

Repository Structure

RiskRadar_AI/
├── data/
│   ├── synthetic/
│   │   └── transactions.csv          # 200 bank transactions (26 fraud)
│   └── policies/
│       ├── High_Value_Transaction_Policy.txt
│       ├── Velocity_and_Burst_Detection_Policy.txt
│       ├── Geographic_Anomaly_Policy.txt
│       └── Device_and_Payment_Instrument_Policy.txt
├── src/
│   ├── fraud/                        # Risk engine, alert models, investigation service
│   ├── graph/                        # Parallel LangGraph workflow and nodes
│   ├── rag/                          # Policy RAG + FAISS
│   ├── tests/                        # Scenario, workflow, alert, and E2E tests
│   └── utils/                        # Data loaders, helpers
├── streamlit_app/                    # Analyst alert queue and investigation UI
├── docs/
│   ├── week2_progress.md
│   └── architecture.md
├── notebooks/                        # EDA and prototyping
├── README.md
├── requirements.txt
├── .env
└── Week1_Progress.md

---

Quick Start

1. Setup

cd /Users/**********/RiskRadar_AI

# Create and activate virtual environment
python3 -m venv .venv
source .venv/bin/activate

# Install dependencies
pip install -r requirements.txt

# Configure environment
# Optional LLM summaries use GOOGLE_API_KEY when USE_LLM_SUMMARY=false

# Keep USE_LLM_SUMMARY=true. If there is enough credits.

2. Explore the Data

python -c "
import pandas as pd
df = pd.read_csv('data/synthetic/transactions.csv')
print('Total transactions:', len(df))
print('Fraud cases:', df['is_fraud'].sum())
print(df[df['is_fraud']==1][['transaction_id','amount','fraud_type']].head())
"

3. Run Streamlit App

streamlit run streamlit_app/app.py

4. Optional Docker Run

RiskRadar AI can also be run as a Docker container for reproducible setup and deployment.

Prerequisite: install Docker Desktop or another Docker-compatible runtime. For older macOS versions such as macOS 12 Monterey, Docker Desktop 4.25.2 was used successfully during local validation.

Build the image:

docker build -t riskradar-ai .

Run the app:

docker run --rm -p 8501:8501 --env-file .env -v "$PWD/data:/app/data" riskradar-ai

Open the Streamlit UI:

http://localhost:8501

The data folder is mounted into the container so analyst feedback logs are persisted on the host machine.

---

Current Capabilities

• Generates fraud alerts from synthetic transaction data.
• Runs a parallel LangGraph investigation workflow.
• Scores risk using triggered policies and weighted rules.
• Retrieves policy evidence with LangChain, local sentence-transformer embeddings, and FAISS.
• Produces risk reasoning, investigation actions, and explainable recommendations.
• Supports optional Gemini-assisted summaries with deterministic fallback.
• Captures human analyst decisions and notes.
• Logs feedback to data/feedback_log.csv for audit and future evaluation.
• Removes handled transactions from the active queue and shows them in Decision History.
• Displays alert metrics, agent trace, policy sources, investigation output, and handled decisions in Streamlit.

---

Key Technologies

• Agent Orchestration: LangGraph
• Optional LLM: Gemini, controlled by USE_LLM_SUMMARY
• RAG Framework: LangChain
• Embeddings: sentence-transformers/all-MiniLM-L6-v2
• Vector Store: FAISS
• UI: Streamlit
• Data: pandas

---

Next Steps

• Add evaluation harness
• Add investigation report export
• Add workflow visualization
• Add historical case retrieval

---

Course: AIOps Capstone (IIT Roorkee)
Theme: 13 — AI Fraud Transaction Investigation Assistant
Project: RiskRadar AI

For setup issues, contact the project team.