MetaFlow – AI Control Plane for OpenMetadata
Winning entry at the OpenMetadata Hackathon (MCP Ecosystem & AI Agents track). Turns OpenMetadata into an autonomous operations platform that detects issues, reasons about impact, and orchestrates resolution across tools without human intervention.
The Challenge
When a data pipeline breaks, resolving it means jumping between OpenMetadata, the warehouse, lineage graphs, and a dozen dashboards — all coordinated by hand. Metadata catalogs know the shape of the data but do nothing when something goes wrong. The knowledge exists; the action doesn't.
- Incident response scattered across disconnected tools
- Manual coordination on every data quality failure
- Metadata catalogs observe but never act
- Root-cause tracing done by hand through lineage
The Approach
Rather than build another dashboard, I turned OpenMetadata into a control plane and put an AI orchestrator on top. The hard call was architecture: a single monolithic agent couldn't reason across detection, lineage, and remediation reliably, so I used a LangGraph supervisor routing to 12 focused specialist agents — each good at one job — with idempotent write-backs so repeated runs never corrupt state.
- OpenMetadata reframed as an action layer, not just a catalog
- LangGraph supervisor pattern over 12 specialist agents
- Idempotent operations safe to retry
- Webhook-driven so it reacts to events, not polling
The Implementation
MetaFlow runs a continuous data steward that detects metadata changes and quality violations and writes health scores back into OpenMetadata's native UI. A contract copilot generates and materializes quality tests, and a self-healing agent classifies failures, traces root causes through lineage, and drafts fixable PR proposals. Fifteen one-click playbooks and webhook auto-triage handle common scenarios end to end.
- Continuous data steward with governance write-back
- Data contract copilot for automated quality tests
- Self-healing remediation with lineage root-cause tracing
- Composite health scores patched into OpenMetadata
- 15 one-click playbooks + webhook auto-triage
The Outcome
MetaFlow won the OpenMetadata Hackathon in the MCP Ecosystem & AI Agents track. It shows what a data catalog becomes when it can act: incidents get detected, reasoned about, and resolved with a proposed fix before a human is ever paged.
- Winner — OpenMetadata Hackathon (MCP & AI Agents track)
- Detection-to-resolution without manual coordination
- Fixes arrive as reviewable PRs, not alerts
- Native integration back into OpenMetadata