The itch
Voice memos are the fastest way to capture a thought and the worst way to find one again. Every recording app produces a graveyard of audio files named by timestamp that nobody ever listens back to. The fix is obvious once you say it out loud: memos should become text you can read, search, and skim.
How it works
While you record, audio streams over a self-hosted LiveKit server to a transcription agent, and Deepgram Nova-2 puts your words on screen as you say them. After you stop, OpenAI Whisper re-transcribes the whole recording with full context, so the version that gets kept is the careful one, not the fast one. GPT-4o-mini then writes a summary, and playback highlights each word as the audio reaches it, thanks to word-level timestamps.
The whole backend is self-hosted: Supabase and LiveKit both run on a VPS behind Caddy. Voice notes are about as personal as data gets, and I would rather they live on a machine I control at a fixed monthly cost.
Status
An early prototype from February 2026. The mobile app works through Expo dev builds (LiveKit’s native modules rule out Expo Go), while the web and desktop apps in the monorepo are labeled phase 2 and phase 3, which is a polite way of saying they are folders. The reasoning behind the two transcription passes is in the case study.