Baxspace

Inspiration

This project came directly from the MCP Hacks Fraud Hunter prompt. We were given the dataset and the review problem, so the goal was not to invent a problem, but to build a tool that solved it well. What shaped our approach was the idea that fraud detection should support a human reviewer, not just output a mysterious score.

How We Built It

We built Fraud Hunter as a Next.js + TypeScript app with a shared detection engine behind both the web UI and the CLI. The engine creates per-card baselines, applies four explainable fraud detectors, and ranks suspicious transactions with plain-English reasons. On top of that, we built a keyboard-first review queue with history, tuning controls, an audit log, and CSV export so a reviewer can move through flagged transactions quickly.

Challenges

The hardest part was balancing detection quality with usability in a short hackathon window. We had to catch multiple fraud patterns, including cross-card merchant bursts, without overwhelming the reviewer with false positives. We also wanted every flag to be explainable, which ruled out a lot of black-box approaches and forced us to make the logic clear and defensible.

What We Learned

We learned that fraud detection is not just a modeling problem, it is a workflow problem. A system is much more useful when it explains why something was flagged and helps a reviewer act on it fast. We also learned the value of keeping the detection logic modular, since separating the engine from the interface made it much easier to iterate on both scoring and UX.

Built With

• next.js-15
• node.js
• react
• tailwind-css
• typescript
• vitest
• zustand