A1 · 2026 · Audio to document
Meeting MOM
Records a client meeting, transcribes it, and returns minutes with action items assigned to the people who were in the room.
Designed it, then built the front end: review flow, keyboard triage, owner correction.
15 to 40 meetings a week at Briqhaus. Nothing goes out until a person signs off on who owes what.
Listen to this one
0:34
Everything that matters gets said out loud, then evaporates
Client meetings ran 4 to 5 hours without reliable records, turning action items into memory arguments and wasting note-takers' participation.
Product Design & Front End: Designed capture, wait, review, and register flows; built front-end triage and keyboard review in Rails.
Enforced 'Model drafts, Human signs off': no action item leaves the building until verified by the meeting chair.
Now default for 15 to 40 meetings a week at Briqhaus; eliminated 33% speaker attribution mismatch errors.
Briqhaus meetings decide budgets, deadlines, and project scope. In 4 to 5 hour sessions, commitments were made verbally and promptly forgotten. Assigning a team member to take notes meant losing an active contributor, and manual summaries still took days to distribute.
I designed the complete experience from blank canvas to production front end in three weeks alongside one backend engineer. The product captures multi-hour audio, structures the discussion into minutes, and extracts action items tied to attendees.
The five-stage pipeline
Everything before the review step is automated and probabilistic. The minutes never leave the building until a human chair signs off on the commitments.
Key product touchpoints
Solving the 33% speaker mismatch bug
Sitting with operators during live client meetings revealed a critical failure: roughly one third of action items named the wrong person. In commercial meetings, misattributing deliverables creates instant embarrassment and erodes trust.
Long meetings were split into 29-minute segments for crash safety. But the transcription API re-indexed speakers from zero on every call. Speaker 0 in chunk 1 was not Speaker 0 in chunk 2, misassigning Priya's commitments to Rahul.
1) Joined segments before calling transcription to maintain consistent voice IDs across the full session. 2) Decoupled action items into discrete database records with one-click owner corrections.
The transcription API numbered voices per call, resetting at the 29-minute seam. Joining chunks before transcription preserved voice continuity.
The model drafts, a person signs off
The model drafts. A person signs off.
Instead of Sending generated summaries directly to clients
Automated summaries that reach clients with the wrong person assigned destroy trust immediately. Our policy enforces that every meeting summary stops at a draft until the meeting chair verifies each item.
After Zero misattributed commitments sent to external clients.
High-speed keyboard triage UI
Reviewing 20 action items is laborious. If the interface is slow, reviewers will rubber-stamp without checking. We engineered a dedicated keyboard-first triage flow.
- Amber state: Items remain amber until a human confirms them.
- Model trace preserved: Original AI suggestions remain visible with a strikethrough, allowing us to track model drift and accuracy over time.
- Auto-advance: Hitting Enter locks the owner and jumps straight to the next item.
Production impact and lessons
Meeting MOM is now the default operating procedure for all client meetings across Briqhaus.
- 15 to 40 meetings logged every week, with minutes distributed within minutes of call conclusion.
- Zero note-takers required, returning active contributors back to leading conversations.
- Lessons learned: Audio diarization in noisy conference rooms will always have an error margin. Designing explicit confirmation workflows is far more effective than trying to prompt-engineer perfection.