MetaGuard AI

Overview

A governance copilot built on top of OpenMetadata that auto-classifies PII, scores metadata quality, and lets you query your data catalog in plain English.

The problem

Data governance in large organisations is manual and inconsistent. Teams tag PII by hand, health metrics live in spreadsheets, and finding the right dataset still means asking someone who has been there long enough to know. I wanted to see how much of that a small AI layer on top of an existing catalog could absorb.

Approach

OpenMetadata is the catalog backbone — it already has an API for everything. On top of it, Gemini 2.5 Flash scans column names and sample metadata for PII signals and writes tags back through the REST API. The health score aggregates completeness signals — descriptions filled, owners set, tags present — into one number. The natural-language agent routes each question to the right endpoint rather than trying to answer from model memory.

What it does

• PII auto-classification using Gemini 2.5 Flash, with tags pushed back through the OpenMetadata APIs
• Governance health score (0–100) derived from metadata completeness and tagging coverage
• Natural-language querying of the catalog via an agent that routes to real endpoints instead of answering from memory
• Built end-to-end during a hackathon sprint

What I learned

Dynamic API routing through an LLM agent is surprisingly reliable once you give it a strict schema of available endpoints. The hallucination rate dropped sharply the moment I switched from free-form descriptions to typed tool definitions.