Garage — AI Across a Workspace Platform

A bot that listens and never publishes

The note-taker joins the LiveKit room as a hidden, subscribe-only bot. It never publishes, so it never appears as a participant with media.

Each participant gets a dedicated Deepgram nova-2 socket, which gives attribution without a diarization model. Transcript segments are persisted as they arrive, so a crash can rehydrate the meeting rather than lose it. The notes go out by email to the resolved participant list.

1 socket per participantsegments persisted on arrival

Single pass or map-reduce, chosen by length

When the transcript fits a 15,000-token budget it gets one gpt-4o pass. Above it, the transcript is split on speaker boundaries, the chunks go through gpt-4o-mini and gpt-4o merges them. The pipeline picks its own shape rather than always paying for map-reduce.

A per-chunk success floor gates the merge: it proceeds only when enough chunks succeeded, rather than silently summarising a hole.

Before any of it, preprocessing strips a known filler set, collapses repeated disfluencies and merges consecutive turns from one speaker. That cuts an estimated 20–40% of input tokens before inference, so it is a cost lever rather than a quality trade.

≤15K tokens → single pass20–40% fewer input tokens