FraudShield AI

• 

  Fraudshield-AI_Report_page-0001

• 

  Fraudshield-AI_Report_page-0002

• 

  Fraudshield-AI_Report_page-0003

• 

  Fraudshield-AI_Report_page-0004

• 

  Fraudshield-AI_Report_page-0005

Inspiration

Online payment fraud costs billions every year, and most victims only find out after the money is gone. We wanted to build an AI agent that evaluates a transaction before it completes and explains its decision, so a human reviewer can trust it instead of facing a black-box score.

What it does

FraudShield AI is a real-time fraud detection API for payment platforms like PayPal. A platform sends it transaction details (amount, time of day, device, location, recent activity), and it returns:

• a risk score from 0 to 1
• a recommended action: approve, review, or block
• human-readable reasons, such as "Unrecognized device" or "Rapid repeated transactions"

Example request:

{"amount": 2500, "hour": 3, "new_device": true, "foreign_location": true, "txns_last_hour": 5}

Example response:

{"risk_score": 0.94, "action": "block", "reasons": ["High transaction amount", "Unrecognized device", "Unusual location", "Rapid repeated transactions", "Late-night activity"]}

How we built it

• FastAPI and Pydantic for the REST API with validated input and auto-generated Swagger docs
• scikit-learn Random Forest classifier that outputs fraud probability
• A rule-based explanation layer that turns risky features into plain-language reasons
• Three-tier decisioning (approve / review / block) so borderline cases go to a human instead of being declined outright
• joblib to persist and load the trained model; Uvicorn to serve it

Challenges we ran into

• Real fraud data is private, so we trained on synthetic transactions that simulate common fraud patterns (large amounts, new devices, foreign locations, rapid-fire transactions, late-night activity).
• Fraud is rare, so the data is heavily imbalanced. We used balanced class weights so the model doesn't just predict "safe" every time.
• Balancing fraud caught against legitimate customers blocked led us to the three-tier threshold design.

What we learned

How to serve an ML model behind a production-style API, how to handle imbalanced classification, and why explainability matters in financial decisions.

What's next for FraudShield AI

• Train on real anonymized datasets (e.g. the Kaggle credit card fraud dataset)
• Integrate with PayPal's sandbox APIs
• Add a reviewer dashboard
• Add authentication, rate limiting, and model monitoring for drift

Built With

• fastapi
• github
• joblib
• machine-learning
• numpy
• pydantic
• python
• rest-api
• scikit-learn
• swagger
• uvicorn