ContextSentinel AI

A DataHub-native AI control plane that prevents unsafe agent data access and autonomously recovers broken data dependencies.

Comment

Inspiration

AI agents are increasingly working directly with enterprise data, but they often lack the metadata context needed to know whether a dataset is safe, governed, fresh, or reliable. A single unsafe request can expose sensitive data, while a failed dataset can silently affect dashboards, pipelines, ML models, and other AI agents.

I built ContextSentinel AI to turn DataHub metadata into something agents can actively reason over and act on.

What it does

ContextSentinel AI has two main agent workflows.

ContextFirewall

ContextFirewall evaluates an AI agent's request before data access is approved. It uses the live DataHub MCP Server to resolve datasets, inspect schema and metadata context, identify sensitive fields, calculate risk, discover safer alternatives, and generate a sanitized approved context.

The decision can then be written back to DataHub so future agents inherit the result.

Phoenix Recovery

Phoenix handles broken data dependencies.

When customer_360_safe fails, Phoenix uses DataHub lineage to discover the downstream blast radius, evaluates replacement datasets, selects customer_360_safe_v2, generates recovery artifacts, and validates the recovery with DuckDB.

The current scenario identifies 16 affected downstream assets and validates the selected recovery with a 93/100 confidence score.

After approval, Phoenix records the recovery decision back into DataHub and verifies the write-back through the MCP Server. The incident state also persists across navigation and browser refreshes.

How I built it

The application uses:

  • DataHub Open Source as the metadata and context platform
  • DataHub MCP Server for live metadata reads and write-back
  • Next.js, React, and TypeScript for the frontend
  • FastAPI and Python for the backend
  • LangGraph for agent workflow orchestration
  • DuckDB for recovery validation
  • Groq for optional LLM-assisted explanations
  • Docker for the local DataHub environment

DataHub is used throughout the actual decision process for dataset resolution, schemas, lineage, governance context, candidate discovery, write-back, and verification.

Challenges

The biggest challenge was making the integration genuinely live instead of presenting static metadata.

This included resolving DataHub entities correctly through MCP, traversing lineage, hydrating real candidate schemas, implementing supported DataHub mutations, verifying write-backs, persisting workflow state, and ensuring LLM rate limits could not break the core agent workflows.

What I learned

I learned that metadata becomes significantly more valuable when agents can use it as operational context rather than simply displaying it to humans.

DataHub allows ContextSentinel to understand what data exists, how it is connected, what policies apply, what will be affected by a failure, and then preserve the result of an agent's action for the next human or AI agent.

What's next

The next step is expanding ContextSentinel from the current recovery and access-control workflows into a broader enterprise agent control plane with additional incident types, automated remediation strategies, and production deployment integrations.

Built With

Share this project:

Updates