Reference: Meeting Assistant

Architectural notes

Reference architecture for a meeting assistant that joins Zoom calls,
transcribes, summarizes, extracts action items, and posts follow-ups.

Why these picks:
- Whisper for transcription quality at the long-form spec.
- Sonnet for summary + action-item extraction; per-meeting eval gates.
- mem0 for per-participant memory ("Jane's open action items").
- Linear MCP creates issues directly from action items.
- Fly.io for long-lived bot processes that hold persistent meeting sessions.

Common gotchas:
- Diarization (who said what) drives summary trustworthiness more than the LLM.
- Action-item false positives erode trust fast; bias toward precision.