ML Lineage Guardian

ML Lineage Guardian architecture

Inspiration

Production ML systems often fail because of silent schema changes, broken lineage, missing ownership, or undocumented dependencies. Teams usually discover these issues only after a pipeline or model has already failed.

What it does

ML Lineage Guardian is a multi-agent system that uses DataHub context to detect risks across the path from datasets and features to ML models and deployments.

The system includes:

Observer Agent - monitors schema, lineage, ownership, and metadata changes
Impact Agent - identifies affected pipelines, features, and production models
Fix Agent - recommends code, configuration, SQL, or migration fixes
Reporter Agent - creates a clear risk report and writes findings back to DataHub or GitHub

How we built it

The project uses DataHub as the metadata and lineage foundation. Python agents access DataHub through its MCP Server or Agent Context Kit, analyze dependencies, classify risk, and generate actionable recommendations.

Challenges

The main challenge is transforming complex metadata and lineage graphs into reliable decisions without producing false alarms. Another challenge is coordinating several agents while preserving context and auditability.

What we learned

We learned how organizational metadata can ground AI agents, reduce hallucinations, and allow them to perform useful work across real data and ML infrastructure.

What's next

Next steps include automated GitHub issues and pull requests, Slack alerts, policy checks, historical change comparison, and support for more ML and orchestration platforms.