A multi-agent protection layer for banks, between the risk engine and payment execution: if someone tries to scam you, it pauses the payment, investigates the receiving account and, if it's a money mule, protects the next victim. Built at Hack the Andes.
Context & problem
The hackathon security track (scams, extortion and impersonation) and the AML alert investigation challenge.
What was built
Non-AI rules detect the risk and hold the money before any agent is called. A coordinator then activates specialists (a Guardian on the customer side and a Detective on the receiving account), a deterministic policy engine decides (allow, verify, pause or escalate), and an Auditor builds a case file with cited evidence for an analyst to confirm. The LLM reasons; it does not decide about the money.
Prototype with 100% synthetic data and 485 automated tests.
Routes considered · Outcome: Case being documented. I'll walk you through it on a call.