Field sales and product teams held long Hindi and English meetings, and follow-ups got lost. We built a recorder app, an agent that transcribes and extracts decisions, risks and action items, and a check that re-finds every quoted line in the transcript. People confirm owners before anything becomes a task.
Before
What we built
A mobile app records for up to three hours, in the background and offline, recovers after a crash, and uploads straight to storage.
A durable workflow transcribes with speaker separation and custom vocabulary, then extracts notes and actions. Long meetings are split, processed in parallel and merged in order.
Code re-finds every quoted line in the transcript within ±90 seconds of its timestamp, Devanagari-safe, and drops citations it can't find.
Summaries cite the moment they came from and play the audio there. Decisions and risks appear as cards with their quote. Action items are edited, ticked and turned into tasks.
Members connect their own coding or chat agent to the workspace. Agents act with the member's access, capped at edit; trash, share and publish always need a person.
Decisions
The model may summarise and extract, but never chooses identity, keys, queries, workflow order or retries. Its only output is a schema.
A ledger of external effects means a paid transcription is never resubmitted when its outcome is uncertain. We crash-tested it at every checkpoint.
A follow-up recording carries the previous meeting's brief and open actions, so the next summary opens with what changed.
The hard part
Notes people can check
Trust came from making every claim checkable: a quote, a timestamp and a replay button, with anything the verifier can't find removed before a person ever sees it.
Measuring it
In testing we measure processing reliability and evidence drop rate. Adoption numbers come after production rollout.
Built with
Next: Production rollout, then answers grounded in approved meeting evidence.
A one-hour call with a founder and an engineer. Bring a BOQ, a stack of POs, last month's invoices or a backlog, and leave knowing what would work, what wouldn't, and what it would take.
170+ projects · 65 companies · founders on every engagement