← All projects
n8nopenaianthropicgeminizoom-api

Zoom lecture QC pipeline

An automated QC pipeline that pulls a finished Zoom recording, cross-checks it with multiple AI models, and only pulls in human review when something's actually flagged.

Full workflow walkthrough available on request.Get in touch

Problem

Reviewing recorded lectures for quality and compliance issues manually is slow and inconsistent — someone has to watch or skim the whole session, and there's no repeatable standard for what actually counts as a real problem versus noise.

Solution

A webhook-triggered n8n pipeline, secured with HMAC signature verification, fires automatically when a Zoom recording is ready. It pulls the meeting's attendance timeline directly from the Zoom API, gathers lecture context from Google Drive, and runs the transcript through parallel analysis passes — including a Gemini-based review and a separate Python-based analysis — before merging everything into one unified timeline. A quality-check AI agent, backed by both OpenAI and Anthropic models with memory and tool access, reviews the merged data and flags anything high-severity. Flagged items get routed through a dedicated verification step before the final report is generated; if nothing's flagged, the report finalizes immediately — so human attention only goes where it's actually needed. The finished report is uploaded to Google Drive, emailed out, and the relevant team members are notified directly.

Architecture

Actual production workflow, built in n8n

Zoom recording ready
   → Webhook (HMAC-verified)
   → Switch (routes the request)
   → in parallel:
        - Zoom OAuth → Extract meeting ID → Get participants → Build attendance timeline
        - Transcript sub-workflow
        - Timeline & chat sub-workflow
        - Gemini-based analysis sub-workflow
        - Python-based analysis sub-workflow
        - Lecture context sub-workflow (Google Drive)
   → Merge (up to 7 parallel branches)
   → Build unified timeline
   → GPT Quality Check (AI agent: OpenAI + Anthropic models, memory, tools)
   → Parse & repair agent output
   → Filter high-severity items
        → needs verification?
             true  → Verify flagged reviews (sub-workflow) → apply verification results
             false → finalize report directly
   → Upload report to Google Drive
   → Email report + notify relevant team members
   → Respond to webhook

Results

Only escalates to a human reviewer when something is actually flagged as high-severity — most sessions get a finalized report automatically, with the verification step reserved for the cases that genuinely need a second look.

Tech stack

n8nZoom APIOpenAIAnthropic (Claude)Google GeminiPythonGoogle DriveGmail

Got a process eating hours every week?

Let's talk