Build idea · Intermediate

An AI meeting-notes summarizer

Turn an hour of rambling into a TL;DR, three decisions, and a list of who owes what.

Upload a recording or paste a transcript and get back a clean, structured summary: a short TL;DR, the decisions that were actually made, and action items tagged with an owner and a due date. The core feature — and the entire point of the build — is structured extraction: you force the model to return typed JSON (decisions[], actionItems[] with assignee) instead of a mushy paragraph, then render it and email it to attendees. Everything else is glue around that one trick.

What you'll learn

  • Forcing reliable, typed output out of an LLM with a schema/tool definition, instead of praying it returns clean JSON you can parse
  • Building an async pipeline: audio upload to transcription webhook to summarization job, so a 90-minute meeting doesn't block an HTTP request
  • Handling long transcripts and context windows — chunking vs. long-context prompting, and what a single meeting actually costs you in tokens
  • The full-stack glue that makes it a product: auth, persisting summaries, and emailing them to attendees

Take it further

  • +Push action items straight into Linear or Notion via their API so nobody has to re-type them
  • +Recurring-meeting memory: embed past summaries in Supabase pgvector and surface 'what changed since last week' for a standing meeting
  • +Auto-ingest: pull recordings from Zoom or a Slack channel and post the finished summary back to the channel automatically