Manuel Porras

All work / Meeting intelligence

Case study · Meeting intelligence

From client meeting to approved task in about 15 minutes

An internal tool of a US marketing agency for home-service businesses. In daily use since April 2026 for more than 30 client meetings a week.

In short: At a US marketing agency for home-service businesses, account managers typed meeting tasks into ProofHub by hand, with up to two days of delay and items getting lost. Now the finished Zoom transcript starts a pipeline: an LLM proposes tasks with likely owners, duplicates are caught, and a person approves or rejects every proposal. Proposed tasks reach Slack and a dashboard in about 15 minutes — in daily use since April 2026 for more than 30 client meetings a week.

30+client meetings per week
~15 minfrom transcript to proposed tasks
2 daysmaximum delay before
100 %of proposals reviewed by a human

Before and after

BeforeNow
The account manager took notes during the meeting.The finished Zoom transcript starts the pipeline automatically.
Afterwards, tasks were created by hand in ProofHub.An LLM proposes tasks with likely owners.
Tedious and error-prone — items got lost.The account manager approves or rejects each one.
Up to 2 days of delay.Results in Slack and a dashboard in about 15 minutes.

How it works

Zoom transcript → approved task Transcript webhook + polling LLM tasks + owners Human approve / reject Project tool task created duplicate? → update existing task ~15 minutes instead of up to 2 days · 30+ meetings a week
Nothing reaches the project tool without a human decision.
  1. Trigger. A Zoom webhook fires when the transcript is ready; polling runs as a safety net in case a webhook is missed.
  2. Understand. An LLM works out who said what and extracts proposed tasks.
  3. Assign. Owners are suggested based on who handled similar ProofHub tasks before.
  4. Between meetings. Requests that clients write in Slack go through the same pipeline.
  5. Review. Proposals appear in Slack and the dashboard; nothing reaches ProofHub without a human click.
About 245 commits since the start of the project.

Four decisions that made it work

1. A human approves every proposal. No proposal lands in ProofHub unreviewed. An auto-push logic with four filters is built and has been piloted.

Why: a wrong task costs more trust than one click costs time — and every decision is data that lets clear cases be automated later.

2. Duplicates are caught. If something was already discussed in an earlier meeting, the existing task is updated instead of creating a new one.

3. Model choice per task, tested on real meetings. Classification stayed on Gemini 2.0 Flash — more accurate and 7× faster in the test. Roadmaps use a newer model that found 4× more items. For meeting scoring, four models were calibrated by consensus and GPT-5.4 was chosen.

4. A guard against silence (09/2026). Recordings without a Zoom transcript are transcribed from audio by Gemini. One 13-minute recording was pure digital silence — and the model invented a complete conversation. Since then, audio below −45 dB is rejected (ffmpeg volumedetect) before it reaches the model.

Why: otherwise the system would have created tasks from a meeting that never happened.

What the review looks like

Proposed taskOwnerAction
Send revised summer-promo budget to client — “…can you send me the new numbers by Thursday?”Dana (account manager), from minute 12Approve · Reject
Check call tracking on the new landing page — “…the calls from that page don’t show up.”Luis (PPC), from minute 27Approve · Reject
Update Google Business Profile hours — “…we’re open Saturdays now.”Dana, from minute 31Merge · Reject — possibly discussed on Oct 6 → update existing task
Illustration with invented test data (fictional client “Acme HVAC”, weekly call, 3 proposals). The real dashboard contains client data and is not shown.

Technology

Node.jsExpressSQLitepm2GeminiGPTZoom S2S OAuthSlack APIProofHub APIDashboard with Google OAuthffmpeg

Source code (cleaned public version with test data): github.com/porrasbrand/zoom-action-items