Manuel Porras

AI Meeting Intelligence Pipeline

Saving teams 10+ hours/week on meeting follow-ups

The Problem

A marketing agency with 30+ weekly client meetings was losing critical action items. Meeting notes were inconsistent, follow-ups fell through the cracks, and the team spent hours manually reviewing recordings. By the time notes were distributed, context was lost and deadlines were missed.

The Approach

I built an end-to-end pipeline that runs autonomously after every Zoom call. The system uses a Zoom Server-to-Server OAuth integration to detect completed recordings, downloads the VTT transcript, and processes it through Gemini 2.0 Flash for structured extraction.

The AI extracts action items with assigned owners, due dates, and priority levels. It identifies decisions made during the call and generates a concise summary. Each extracted item includes a verbatim transcript excerpt so the team can verify context without re-watching the recording.

Results are posted to Slack within 15 minutes of a call ending, and action items can be pushed directly to ProofHub for project tracking. The system includes confidence scoring, adversarial verification to catch hallucinated items, and a web dashboard for review and editing.

The Outcome
30+
Client meetings processed weekly
15 min
Turnaround (was 2-day delay)
10+
Hours saved per week
Tech Stack
Node.jsGemini 2.0 FlashZoom S2S APISlack APISQLiteVTT ParsingExpressProofHub API
Related

Blog: Building an Automated Meeting Action Items Pipeline

← Back to all projects