The aim was that a commitment made on a call could be found and proven later. Here is how the shipped pipeline differs from what came before it.


Sat in on a week of calls across two teams. The real gap was decisions nobody could later prove were made.
Built ingestion for live calls joined through the conferencing tool and for uploaded recordings, converging on one transcription step.
Clean demo audio falls apart on a real three-person call with a dog barking. We trained against real recordings, not studio ones.
An early version flagged nearly every sentence as an action item. The shipped version only tags explicit commitments.
Tested the sync and search layer against a real backlog of past calls before anyone relied on it.
The second pilot team joined only after the first team’s edge cases were handled; a month of real use surfaced what scripts never would.
