Mule Interceptor
An autonomous money-mule detection and response agent. It reads the flow of money as a live graph, interrupts the cash-out, and drafts the report.
The problem
Banks do three jobs slowly and separately: they detect a money-mule network, they intervene, and they report it. By the time the manual chain catches up, the funds are gone. Built for the Zenith Bank Zecathon 6.0 fraud and financial-crime track, Mule Interceptor collapses those three jobs into one automatic flow.
How it works
- Graph Sentinel fuses explicit AML rules, graph features over the transaction network (fan-in collectors, short cash-out paths), and an unsupervised anomaly model into a single risk score that arms an account.
- Active Harness holds the next real cash-out from an armed account and requires a step-up check. A genuine customer clears it; a mule farm fails, and the funds are frozen before the money leaves.
- The agent assembles the evidence and drafts a Suspicious Activity Report for a compliance officer to review and file.
The design decision that matters
The language model writes prose only. A deterministic state machine owns every decision and all control flow, so the same case always produces the same actions and the same document shape. That is what makes the agent auditable, which is the whole point in a compliance setting. The freeze is automatic; the filing to the regulator is always a human step.