Zeta

Meetings on Google Meet and Microsoft Teams leave nothing behind but memory and whoever volunteered to take notes. Transcripts, when they exist, arrive after the fact and never say who said what. Summaries are a manual chore, and the follow-up email is the first thing to slip.

Zeta had to make the record automatic: capture the call as it happens, know who is speaking, produce a summary worth reading, and put it in people's inboxes without a human in the loop.

A Chrome extension captures the meeting from inside Meet or Teams and streams audio to the backend over WebSockets, so the transcript builds in real time instead of after upload. Transcription and speaker diarization run in workers behind a BullMQ queue on Redis, which keeps the live path responsive while the heavy processing scales independently.

Once a meeting ends, Gemini turns the diarized transcript into a structured summary. Supabase holds auth, meetings, and collections, calendar integration attaches context to each session, and the summary goes out by email automatically. The whole thing runs in Docker on AWS.

Zeta went from empty repo to a working production platform in six weeks: live transcription over WebSockets, diarization, Gemini summaries, automated email distribution, calendar integration, collections, and the extension that makes capture one click.

Putting the transcription work behind a queue from the first week is what kept the live path stable while everything else was still moving. Next time I would design worker backpressure and retries on day one instead of adding them when the first long meeting exposed the gap.