NEXORA

NEXORA

Inspiration

Most AI agents answer with confidence even when they're guessing. We wanted to build the opposite: an agent that only answers from evidence it can actually point to — and when it can't find that evidence, it says so instead of making something up.

NEXORA started as a company knowledge graph — connecting people, projects, documents, decisions and meetings so an AI agent could answer questions like "why is this project delayed?" with real sources, not a plausible-sounding guess. When we saw the DataHub Agent Hackathon, the fit was obvious: DataHub already solves this exact problem for company data — tables, pipelines, dashboards, lineage — so we extended NEXORA with a "Data Assets" agent that applies the same principle to a live DataHub instance via its MCP Server.

What it does

An interactive knowledge graph of company entities (people, projects, decisions, documents, meetings) with an AI agent that answers questions using real graph evidence and cites its sources
A "Data Assets" agent that connects to a live DataHub instance through DataHub's MCP Server, discovers whatever tools that instance actually exposes, and answers questions about your real data estate
If DataHub isn't configured or reachable, the agent says so plainly — available: false with a clear reason — rather than fabricating an answer. Same rule applies everywhere else in the app: our "search any company" feature only has real, sourced data for a small hand-verified set of companies, and is honest about not having more.

How we built it

React + TypeScript + Vite on the frontend, FastAPI + SQLAlchemy on the backend. The DataHub integration uses the official mcp Python SDK — our client discovers DataHub's available tools at connect time rather than hardcoding tool names that can change between DataHub versions, then calls whichever tool looks like a search/discovery tool with the user's question.

Challenges we ran into

We didn't have a running DataHub instance available while building the integration (Docker/WSL setup on Windows ate a real chunk of our time). Rather than ship untested code, we built a small fake MCP server that mimics DataHub's shape and wrote automated tests that run our actual client against it over a real MCP session — proving the connection, tool-discovery, and tool-calling logic genuinely works, even though we couldn't verify DataHub's exact tool schema until Docker was finally working.

What we learned

Being upfront about what's verified versus what's assumed made for a better product, not a weaker one — an agent (or a hackathon submission) that clearly states its limits is more trustworthy than one that quietly guesses.