Recap pipeline
The sequence that produces a meeting recap once the meeting ends: transcript ingestion, action-item extraction, owner attribution, channel routing, delivery.
Why it matters
A meeting recap is more than a transcript summary. The good version: short (5-8 bullets), organized (decisions, action items by owner, follow-ups), action-oriented (with drafts attached for the obvious next steps). Producing this from a 90-minute transcript is a multi-step pipeline; getting any step wrong damages the user's trust.
The harder problem is doing it cheaply at scale. A naive recap pipeline that calls a frontier model on every transcript costs more than the value it produces. The right pipeline uses smaller models for high-volume work and larger models only for synthesis.
How Ajeris handles this
The recap pipeline runs in three phases: load context, run the external work (transcript fetch, model calls, tier classification), then persist results. Action items are tier-classified (auto, approval, always-ask) and routed to the user's preferred channels. The pipeline aims for end-of-meeting to recap-in-inbox in under five minutes for typical meetings.
