Fraud Detection MCP

AI-powered fraud detection and investigation platform using a real financial fraud dataset, SQLite in-memory SQL, MLflow model tracking, MCP tools, and LangGraph agents.

The project is intentionally built like a production-style AI/ML system rather than a notebook:

• Real dataset: Bank Account Fraud Dataset Suite from Kaggle / NeurIPS 2022.

• SQL layer: CSV is loaded into SQLite in-memory, so all fraud tools query SQL tables.

• MLflow: model training logs metrics, parameters, artifacts, and the trained model.

• OpenAI: investigation summaries require OPENAI_API_KEY from .env.

• LangGraph: orchestrates the fraud investigation workflow.

• MCP: exposes fraud tools as a Model Context Protocol server.

• FastAPI: serves investigation endpoints and a small dashboard page.

Architecture

BAF Kaggle Dataset
  ↓
scripts/download_dataset.py
  ↓
data/processed/baf_base_sample.csv
  ↓
scripts/train_model.py
  ↓
MLflow run + models/fraud_model.joblib
  ↓
FastAPI app
  ↓
SQLite :memory: SQL database
  ↓
Fraud tools
  ↓
LangGraph Fraud Agent
  ↓
OpenAI investigation summary

MCP Server exposes the same fraud tools to MCP-compatible clients.

Related MCP server: DevFlow MCP

1. Create and activate venv

Windows:

py -3.11 -m venv .venv
.venv\Scripts\activate

Mac / Linux:

python3 -m venv .venv
source .venv/bin/activate

Install dependencies:

pip install -r requirements.txt

2. Configure OpenAI key

Create .env from the example:

copy .env.example .env

Mac / Linux:

cp .env.example .env

Then edit .env:

OPENAI_API_KEY=sk-your-real-key
OPENAI_MODEL=gpt-4.1-mini

Do not commit .env to GitHub.

3. Download the real BAF dataset

This uses KaggleHub, not a fake dataset.

python scripts/download_dataset.py

It downloads:

sgpjesus/bank-account-fraud-dataset-neurips-2022

Then it copies the base CSV and creates:

data/processed/baf_base_sample.csv

By default it samples 50,000 rows for fast local development. You can change --sample-size.

Example:

python scripts/download_dataset.py --sample-size 100000

4. Train the model with MLflow

python scripts/train_model.py

Outputs:

models/fraud_model.joblib
models/fraud_model_metadata.json
mlruns/

Open MLflow UI:

mlflow ui --backend-store-uri ./mlruns

Then open:

http://127.0.0.1:5000

5. Run the FastAPI app

uvicorn app.main:app --reload

Open:

http://127.0.0.1:8000

Useful endpoints:

GET  /health
GET  /applications/high-risk?limit=10
GET  /applications/{application_id}
POST /applications/{application_id}/score
POST /applications/{application_id}/investigate
GET  /cases
GET  /audit

Example:

curl -X POST http://127.0.0.1:8000/applications/10/investigate

6. Run the MCP server

python -m mcp_servers.fraud_mcp_server

The MCP server exposes tools such as:

get_application
list_high_risk_applications
score_application
investigate_application
create_review_case
get_review_cases
get_audit_log
safe_select_query

LangGraph workflow

START
  ↓
load_application
  ↓
score_application
  ↓
policy_decision
  ↓
llm_investigation
  ↓
maybe_create_review_case
  ↓
END

The agent uses deterministic tools for data access and scoring, then calls OpenAI to produce an analyst-style investigation summary based on evidence.

Why SQLite in-memory?

The first version uses SQLite :memory: to provide a real SQL query layer without requiring PostgreSQL or Docker. The storage layer is abstracted so it can later be replaced with PostgreSQL while keeping the MCP tools and LangGraph agent almost unchanged.

Suggested GitHub description

Fraud detection and investigation platform using OpenAI, LangGraph, MCP tools, MLflow, SQLite in-memory, and the Bank Account Fraud Dataset.

Suggested topics:

fraud-detection, mcp, ai-agents, langgraph, openai, mlflow, sqlite, fastapi, machine-learning, financial-ai