Inspiration
Built from real fraud-defense needs: quick rule checks plus behavior-based detection. We wanted a clean demo that shows how rule logic, sequence model scoring, and human review can work together in real time.
What it does
SignalShield scores each transaction with a fused risk signal (rule + transformer model), returns risk_score, confidence bounds, and a decision (allow, review, block). Medium-risk events create a review ticket for an analyst console.
How we built it
FastAPI backend for APIs, in-memory user/transaction stores, optional transformer model for sequence fraud scoring, static browser UIs for user and analyst flows. The system can run locally with uvicorn and simple HTTP file serving.
Challenges we ran into
Imbalanced fraud labels, sequence-length and timestamp normalization, and making sure the demo still works when model artifacts are missing.
Accomplishments that we're proud of
A complete end-to-end demo with hybrid scoring, confidence interval output, reasons and explainability, plus human-in-the-loop review queue in a compact repo.
What we learned
Real risk systems need layered defense, graceful degradation (model optional), and analyst workflows for borderline cases.
What's next for SignalShield - Real Time Fraud Detection
Add persistent storage (Postgres/Redis), JWT auth, event stream processing, richer feature pipelines and scalable deployment (containers + CI/CD).
Built With
- css
- fastapi
- html
- javascript
- python
- pytorch
- transformer
- uvicorn
Log in or sign up for Devpost to join the conversation.