Fraud Decision Engine

A machine learning project that converts fraud-risk predictions into operational business decisions using a cost-aware 3-tier fraud policy.

Live Demo

• App: https://fraud-decision-engine.onrender.com/app
• API Docs: https://fraud-decision-engine.onrender.com/docs

---

Overview

Most fraud-detection projects stop at classification metrics like accuracy or ROC-AUC. This project focuses on what happens after prediction — deciding whether to allow, review, or block a transaction based on business risk and operational cost.

The system applies a 3-tier decision strategy:

Decision	Purpose
ALLOW	Low-risk transactions
REVIEW	Medium-risk transactions sent for analyst review
BLOCK	High-risk transactions

Using optimized review and block thresholds, the policy reduced simulated fraud decision cost from approximately 4,800 to 2,996 (~37.6%) compared to a naive threshold baseline under predefined fraud-loss and operational-cost assumptions.

---

Key Features

• Real-time fraud scoring with FastAPI
• Batch transaction inference
• Threshold-based policy simulation
• Cost-impact evaluation workflows
• Interactive frontend for fraud-risk testing
• API validation and structured inference pipelines
• CI testing with GitHub Actions
• Public cloud deployment on Render

---

Tech Stack

Backend: FastAPI, Python, Scikit-learn Frontend: HTML, CSS, JavaScript Testing & DevOps: Pytest, Docker, GitHub Actions Deployment: Render

---

Run Locally

Clone the repository:

git clone https://github.com/sreenidhik31/fraud-decision-engine.git
cd fraud-decision-engine

Create and activate a virtual environment:

python -m venv .venv
.venv\Scripts\activate

Install dependencies:

pip install -r requirements.txt

Start the FastAPI application:

uvicorn api:app --reload

Open locally:

• App: http://127.0.0.1:8000/app
• API Docs: http://127.0.0.1:8000/docs
• Policy Metadata: http://127.0.0.1:8000/policy

---

Run Tests

pytest tests/ -v

---

Run with Docker

Build the container:

docker build -t fraud-decision-engine .

Run the container:

docker run -p 8000:8000 fraud-decision-engine

Open locally:

• App: http://127.0.0.1:8000/app
• API Docs: http://127.0.0.1:8000/docs

---

Repository Structure

fraud-decision-engine/
├── .github/
│   └── workflows/
├── model/
├── notebooks/
├── static/
├── templates/
├── tests/
├── api.py
├── scoring.py
├── requirements.txt
└── Dockerfile

Directory Notes

• model/ Serialized Scikit-learn model artifacts and threshold metadata.

• notebooks/ Threshold tuning, policy simulation, and offline evaluation workflows.

• tests/ API validation and inference pipeline tests.

• static/ and templates/ Frontend assets and HTML templates served directly from FastAPI.

---

API Endpoints

Method	Endpoint	Description
POST	/score	Single transaction scoring
POST	/score-batch	Batch transaction inference
POST	/simulate-policy	Threshold-based policy simulation
POST	/evaluate-policy-batch	Batch policy evaluation
POST	/simulate-cost-impact	Compare policy cost impact
GET	/policy	Governance and policy metadata
GET	/metrics	API usage metrics
GET	/model-info	Model metadata and schema
GET	/ready	Readiness check

---

Highlights

• Designed a cost-sensitive fraud decision workflow instead of a standard binary classifier
• Reduced simulated fraud decision cost by ~37.6% using optimized threshold policies
• Built REST APIs for scoring, simulation, and batch evaluation
• Implemented schema validation, feature alignment, and model caching
• Deployed a live public application with CI testing workflows
• Structured the project using modular inference and policy layers

---

Future Improvements

• Real-time streaming inference
• Drift monitoring and retraining workflows
• Analyst feedback integration
• Production-grade monitoring and observability
• Feature-store integration