Overview
A governance agent for data catalogs in which deterministic rules identify what is wrong and a language model only explains the findings and proposes fixes.
The problem
Data catalogs steadily accumulate governance problems: datasets without owners, missing documentation, inconsistent tags and classification gaps.
Generic AI chatbots asked to “review” metadata produce confident answers that cannot be verified or reproduced — a poor fit for governance work.
The solution
DataGuardian AI separates judgement from explanation. Deterministic rule checks run against DataHub metadata (read-only) to decide what is wrong; an agent workflow then explains each finding in plain language and suggests remediation.
Findings are presented in a React workspace, and a zero-setup demo mode runs on a deterministic 25-asset catalog so anyone can evaluate it without credentials.
Engineering highlights
Rules decide, AI explains
Findings are reproducible because the language model never decides what is wrong.
Agent workflow with LangGraph
A structured multi-step workflow rather than a free-form chat loop.
Read-only by design
The agent reads catalog metadata and never writes to DataHub.
Quality gates in CI
GitHub Actions runs the automated test suite, mypy type checks and ruff/oxlint linting on every change.
What was built
- Rule engine for metadata governance checks
- LangGraph agent for explanations and fixes
- FastAPI backend with DataHub integration
- React and TypeScript review workspace
- Deterministic demo mode
- CI pipeline with tests, type checks and linting
Results
We only publish results we can verify. This project has no published performance or business metrics — we would rather show the engineering than estimate impact.