"Inspiration"
Data catalogs store metadata but don’t act on it. We wanted DataHub to work, not just describe.
"What it does" Four AI agents read DataHub, do real work (enrich docs, generate SQL, monitor ML, auto-tag PII), and write results back. A chat interface (text + voice) commands them naturally.
"How we built it" FastAPI + LangGraph + Gemini + DataHub Python SDK. Frontend: Tailwind + D3.js + Web Speech API. LangSmith traces every call.
"Challenges" Secret scanning blocked pushes. Fixed with .gitignore and orphan branches. Gemini function-calling needed a keyword fallback. Scroll-free layout required precise CSS.
"Accomplishments" All 4 agents work in mock mode. Cinematic stream-to-chat animation. Chat answers any DataHub question. LangSmith auditable. Runs with one command.
"What we learned"
Function-calling turns a bot into an agent. Mock-first equals instant demos. Animation wins judges. Trace everything.
"What’s next" Real DataHub writes, multi-agent teamwork, Airflow/Dagster codegen, Slack integration, MCP Server support.
Log in or sign up for Devpost to join the conversation.