meeting-autopiolt

The meeting assistant that takes the notes for you.

Comment

Inspiration

Every meeting ends with messy notes, forgotten decisions, and action items that vanish into someone's inbox. We built Meeting Autopilot so that no decision gets lost — software should sit in your meeting quietly, capture who said what, and hand you the summary and tasks when you're done.

What it does

Meeting Autopilot joins your meeting like a personal assistant. It captions every speaker live with color-coded, speaker-aware labels using Deepgram's diarization. When the meeting ends, one click produces a clean summary, a list of action items each assigned to the right person, and a ready-to-send email recapping decisions and next steps. It works three ways: a built-in demo (zero setup), an uploaded recording, or a live microphone.

How we built it

React + Vite frontend with a full-screen reel interface. A Node.js + Express backend with a WebSocket endpoint streams mic audio to Deepgram (nova-3 model with speaker diarization) and returns live transcript events. Groq's LLM (openai/gpt-oss-20b) turns the transcript into structured notes — summary, topics, action items with owners, and an email draft. Deployed as a single unified service on Render.

Challenges we ran into

Debugging live audio streaming end-to-end. The Deepgram SDK crashed on Node 24 ("Invalid Sec-WebSocket-Protocol"), so we switched to a direct WebSocket connection. Then WebSocket frames arrived as binary Buffers instead of strings, silently dropping our config messages — fixed by checking the isBinary flag. Finally, the audio encoding had to be declared explicitly (linear16, 16kHz) or Deepgram closes the stream. Each issue surfaced only through careful logging and testing on the live server.

Accomplishments that we're proud of

A fully working live meeting flow where multiple speakers are separated automatically. A clean, demo-friendly reel UI that needs no login, no setup, and works on a plain laptop mic. And building the whole project — design, backend, AI integration, deployment — end-to-end.

What we learned

Streaming AI services behave differently from REST APIs — connection setup, subprotocols, and binary vs text frames all matter. We also learned how to debug deployed services: adding logging, test scripts that mimic the browser, and reading close codes instead of guessing.

What's next for Meeting Autopilot

Add calendar integrations to auto-join scheduled meetings, support more languages, surface topic flags for follow-ups, and a review screen where you can edit and approve the notes before they're sent.

Built With

Share this project:

Updates

Submission history