Work

Meet

Video calls on my own LiveKit, with the transcript and the decisions written up after.

Internal tool
kossflow.com
Meet

Problem

After a call nobody remembers who promised what, and the recording nobody rewatches is not a record.

Solution

A self-hosted video-meeting platform on LiveKit with a voice AI assistant inside the call - wake-word gated, answering over the organisation knowledge base. Transcription is speaker-attributed, with an STT fallback chain: Groq Whisper, then Deepgram, then a self-hosted model.

How it is used

  1. 01Open a room and talk - the assistant only wakes on its word.
  2. 02Ask it something mid-call; it answers from the organisation knowledge base.
  3. 03Take the transcript afterwards, and the meeting turned into a specification PDF.

Results

  • Three-step STT fallback, so a provider outage does not lose the call.
  • Meeting-to-specification PDF produced automatically.
Access
Internal tool
Kind
Real-time

Stack

  • Next.js 16
  • LiveKit
  • Whisper
  • Postgres

Want to talk about it?