Every new user's platform question used to pull in a PM and the engineers who built that piece. As a side project, I built an assistant over the platform's documentation, connected through MCP to Confluence, Databricks, Azure, AWS and Collibra, and every answer links the document it came from. Questions about our own deployment are answered only from internal sources; general questions can fall back to web search. I wrote its algorithm with another PM and key engineers.

I designed its two-stage evaluation on my own: a retrieval check against known answers, then a rubric for answer quality, fidelity to source, helpfulness, structure and whether it resolved the user's purpose.

Result. 200 users, reached through word of mouth and a requirement to ask it before escalating to engineering.