Meetings shouldn't be where action items go to die

Why we built this

Every team has the same leak. A decision is made on a call, someone agrees to own it, a deadline is floated, a concern is raised and half-answered — and then the meeting ends. What survives is a transcript nobody reopens and a to-do list that was never written down. The tools meant to help stop at transcription and summarisation. They do not weigh urgency, they do not notice when a discussion turned tense, and they never open a ticket.

We wanted the opposite: a system that reads the conversation the way a good chief of staff would — catching not just what was said but how it landed — drafts the follow-up work, and then, once a person signs off, actually creates it in the tools the team already uses.

What Prosody does

You upload a meeting transcript exported from Zoom or Microsoft Teams. Prosody parses it into speaker segments and scores every line on three emotional dimensions — Valence, Arousal and Dominance — using a RoBERTa model fine-tuned on the EmoBank dataset. That gives you a mood timeline and shows exactly where a discussion got tense.

Claude then reads the transcript and the emotion scores together and drafts action items, each with a suggested assignee, an AI confidence score, a priority, and a full ticket description, issue type, story-point estimate, due date and labels. A human reviews every draft, confirms the real Jira or ClickUp account for the assignee, edits anything, or rejects it. On confirm, Prosody creates the issue in the company's Jira or ClickUp — asynchronously, with retries and per-tenant circuit breakers — and drops it into the right sprint under the right epic.

Every confirm, edit and reject is recorded as feedback, powering a company-wide analytics dashboard: completion trends, workload distribution, recurring tension patterns and prediction accuracy. Transcript content is only ever sent to the Claude API; OAuth tokens are encrypted at rest, and every tenant's data is strictly isolated.