Governed AI Metadata Assistant

A government data office needed to raise metadata quality across their data lakehouse estate — reviewing it by hand doesn’t scale, and letting a model rewrite metadata directly was never going to pass governance. I designed a three-agent architecture: an Extraction Agent that turns the client’s own standards documents into structured knowledge; an Assessment Agent that scores each dataset field and cites the source for every recommendation; and an Improvement Agent that turns reviewer feedback into candidate rules an admin approves.

The discovery questions shaped the architecture more than anything else: “When the system finds a problem, what happens next?” — the answer (a human decides) made human-in-the-loop the architecture, not a bolted-on feature. “Can data leave your environment?” — no, which forced deployment inside the client’s perimeter with model access routed through their own gateway. “If two teams use this, can one see the other’s data?” — no, which made feed isolation a hard security boundary rather than a convenience.

The system is advisory only, by architecture — it has no write path to the actual catalogue. Ungrounded recommendations (no traceable source) are rejected before a human ever sees them. That structural constraint, more than any policy, is what made the system approvable in a regulated environment.