Planning KubeCon usually means scrolling Sched and scanning abstracts one at a time, keyword search misses synonyms and tells you nothing about what's trending.

An MCP server sits between an MCP-compatible AI client or chat bot and a semantic index over every KubeCon session, abstracts video transcripts, enriched with speaker, company, and tool metadata. Four tools (search, ask, lookup, details) power session retrieval and conversational queries like "Can you tell me more about session X", "which previous KubeCons covered Y subject?", or "build me an agenda focused on domain Z."

Most of the design work happens on the corpus side: metadata narrows candidates before vector search, and pre-built vector snapshots cut cold start from 30s to <1s. We iterate through video transcript sessions, chunking each talk (~4k words vs 200 in abstracts) and storing them in object storage so the system can answer deeper, content-level questions abstracts alone can't.