Multi-agent compliance intelligence
A compliance system should be able to explain every material finding, show its regulatory source, and preserve a reviewable trail instead of asking a human to trust an opaque score.
Independent system · architecture, implementation, evidence, and delivery
ARGUS was selected as one of three Microsoft Agents League 2026 Hack for Good winners, and its architecture was published as a guest post on Microsoft Tech Community.
Manual KYC/AML review doesn't scale, and unaudited AI decisions don't survive a regulator's audit.
An orchestrator dispatches identity, screening, corporate intelligence and transaction intelligence over A2A in parallel; their four results fan in to a single compliance and risk agent that scores them. Five specialists, each with its own tools, rather than one general prompt asked to be all of them.
Foundry IQ is wired to exactly two of the five agents: screening queries the sanctions and adverse-media knowledge bases, compliance queries the regulations base. The agents that need no cited knowledge do not get a retrieval path they could misuse.
The result exposes agent participation, tool activity, confidence, drivers, citations, and recommended actions so a reviewer can inspect how the report was assembled.
ARGUS is a technology demonstration using synthetic data. It is not a licensed compliance product and must not make real KYC/AML decisions.
I’m always interested in the constraints that make an architecture worth explaining.