Hackathon · 2026 · Cursor × Briefcase · London 2026
Atlas
A control room that watches a company's payments in real time and sends AI agents to investigate the suspicious ones.
Python · FastAPI · Next.js · Claude API
The problem
Finance teams see thousands of transactions a day. The genuinely risky ones hide in the noise, and no single rule or model catches them all.
What I built
I built a pipeline where simple rules and a fast model triage every transaction, a stronger model analyses the unclear ones, and the hardest cases spawn a research agent that writes a memo for the human reviewer.
What it does
- Cheap Checks First. Simple rules and a fast model handle the flood of transactions, so the expensive model only sees the ones worth a closer look.
- The Company on a Globe. Every counterparty is plotted on a live globe with its health status, pulsing as new transactions come in.
- Agents That Investigate. The hardest cases spawn a background agent that digs through public records and web sources, then drops a memo for the reviewer.
- Never Moves the Money. Atlas only recommends. Every decision ends with a person in finance, not the machine.
Gallery
The full engineering is on GitHub.