Inspiration
I was inspired by real enterprise investigations where answering one question meant reading multiple requirements documents, tracing data across functional areas, and manually building chronological SQL. A “small” schema change can affect finance, payments, returns, dashboards, and stored procedures, but teams rarely have one trusted place showing the complete impact.
What it does
contextIsKey turns DataHub metadata into an evidence-backed investigation and fix plan. Users can upload an SRS or incident document, map each requirement to governed assets and columns, generate cross-domain investigation SQL, trace lineage, discover hidden stored-procedure dependencies, visualize regional exposure, and produce reviewable fixes with validation and rollback steps.
Every recommendation shows the DataHub context that supported it. Unsupported rules remain visibly unmapped instead of being invented, SQL is never automatically executed, and metadata write-backs require human approval.
How we built it
I built contextIsKey with Python 3.12, FastAPI, Jinja2, SQLGlot, OpenAI, and the official DataHub MCP interface. The agent retrieves asset search results, schemas, lineage, ownership, glossary context, entities, and query history before generating an answer.
A tenant-ready MCP adapter supports live DataHub context, while DataHub GraphQL handles approval-gated write-back proposals. The hosted Railway demo uses a deterministic synthetic DataHub-shaped enterprise graph so the complete workflow remains reliable and transparent without claiming a live cloud tenant.
Challenges we ran into
The hardest challenge was preventing the AI from producing convincing but unsupported answers. I added strict grounding rules so OpenAI can organize requirements but cannot invent DataHub assets, columns, owners, lineage edges, or execution results.
I also could not access a DataHub Cloud trial using a personal email. Instead of hiding that limitation, I created a deterministic hosted mode and verified the live integration locally through the official DataHub MCP server.
Another challenge was combining metadata lineage with dependencies that may exist inside stored procedures, views, casts, and dynamic SQL. contextIsKey handles this by combining DataHub’s graph with a generated read-only database discovery query.
Accomplishments that we're proud of
I’m proud that DataHub is the foundation of the workflow rather than a feature added for presentation. The product repeatedly uses DataHub context to constrain document mapping, SQL generation, impact analysis, regional risk, ownership coordination, and proposed fixes.
The demo includes:
- SRS and incident-document ingestion
- Cross-domain chronological SQL generation
- A visible, numbered DataHub evidence trail
- Column-level impact and hidden-consumer discovery
- Regional exposure and owner-coordination dashboards
- Reviewable fixes, validation, rollback, JSON, PDF, TXT, and SARIF exports
- Approval-gated DataHub action proposals
- 195 passing automated tests and a deployed Railway application
What we learned
I learned that DataHub is fundamentally different from a normal database. A database stores rows; DataHub explains what the data means, where it came from, where it flows, who owns it, and whether it can be trusted.
That context is what makes enterprise AI useful. A model can generate SQL alone, but it needs governed metadata to generate SQL that is explainable, reviewable, and grounded in the organization’s actual systems.
What's next for contextIsKey
Next, I would connect a production DataHub Cloud tenant and incorporate live quality, freshness, usage, dashboard, and ML-model metadata. Real-time metadata events could automatically trigger impact investigations whenever schemas or quality signals change.
I would also add more warehouse dialects, SSO and role-based approvals, persistent investigation history, automated evaluation against resolved incidents, pull-request generation for approved fixes, and fully audited DataHub write-back workflows.
Log in or sign up for Devpost to join the conversation.