What they built

CloseLoop AI - Meeting2Mission

Most meeting assistants stop at notes. CloseLoop AI continues after the meeting: it reads a transcript, extracts the real decisions (not just a summary), figures out who owns each one and by when, creates the Jira ticket, schedules the calendar deadline, posts the Slack follow-up, and later chases anything that goes overdue — with zero manual handoff between "we decided X" and "it's assigned, scheduled, and being followed up on." The Problem Meeting transcripts contain names and natural-language deadlines ("Sarah will finish the dashboard by Friday") — never email addresses, Jira project keys, Slack channel IDs, or ISO timestamps. Every meeting-assistant tool we tried either asked the user questions a transcript can't answer, or silently guessed and got it wrong. We treated that as a schema-design problem, not a prompting problem, and fixed it at the tool-contract level. How It Works Meeting → Transcript → Decision Extraction → Owner Resolution → Deadline Resolution → Jira Ticket → Calendar Deadline → Slack Follow-up → Escalation → Weekly Report (Notion + Email) → Executive Dashboard Real structured extraction via Gemini — decisions, owners, deadlines, priority, action items, risks, with a confidence score — not a vague summary. A human-in-the-loop confidence gate: anything under 80% confidence never becomes a ticket automatically; it's routed to a Slack approval request instead. An identity resolver that turns "Sarah Chen" into her real email/Slack/Jira/GitHub identity, backed by a small learned team-member directory, with an explicit ambiguous/not-found state instead of a silent wrong guess. A deadline resolver that turns "Friday", "next week", "end of sprint" into an ISO date relative to when the meeting actually happened. One-call orchestration (process_meeting) that runs the entire pipeline server-side instead of forcing the AI to manually chain 5+ tool calls.