MeetSummary AI

MeetSummary turns every client call into a structured record — speaker-labelled transcript, AI summary, action items, and a ready-to-send follow-up. Automatically.

Comment

MeetSummary: From Overwhelmed VA to Building My Own Meeting Intelligence Tool

Inspiration

Virtual Assistants spend 2–4 hours every week just processing meetings — transcribing, summarising, writing follow-up emails, updating CRMs. That's billable time spent on admin, not strategy. I built MeetSummary to eliminate that entirely: one upload, everything handled.

What it does

MeetSummary is a meeting intelligence platform for Virtual Assistants and client-facing professionals. Upload any audio or video recording — or start a live session directly in the browser — and within minutes you get a full speaker-labelled transcript, an AI-generated executive summary, extracted action items with owners and due dates, a ready-to-send follow-up email, and direct sync to HubSpot, Salesforce, Slack, or Teams. Nothing leaves the workflow manually.

How I built it

Deepgram Nova-3 handles real-time transcription and speaker diarization over a WebSocket stream. OpenRouter routes AI tasks (summaries, action item extraction, live coaching, translation) through a free-model fallback chain so the app works without a paid LLM subscription. Supabase powers auth, the database, realtime subscriptions, and edge functions. Cloudflare R2 stores audio files. The frontend is React 18 + Tailwind CSS v4, built in Figma Make.

The entire project was built while still handling and learning VA skills daily.

Challenges I ran into

Getting live speaker diarization right was the hardest part. Deepgram's v1 streaming diarizer sometimes reassigns speaker IDs mid-session, creating phantom third speakers in two-person calls. We built a session-scoped speaker normalisation layer that detects ID drift and remaps it to the correct existing speaker using a recency window. The other challenge: building a complete post-meeting workflow — structured JSON summaries, PDF-exportable minutes, CRM sync, email drafting — that degrades gracefully when API keys are missing or models fail, without ever crashing or blocking the user.

Accomplishments that I'm proud of

A fully working end-to-end VA workflow: live transcription → AI summary → action items → follow-up email → CRM push → Slack/Teams notification, all from one interface. Plus speaker diarization, compliance/PII redaction mode, analytics dashboard, recurring meeting intelligence, and a formal meeting minutes view with PDF export.

What I learned

Free LLM models on OpenRouter are capable enough for meeting summarisation — but require a resilient fallback chain because availability changes daily. Deepgram's streaming v1 diarizer is reliable but needs client-side normalisation to be production-quality. Supabase Realtime with postgres_changes is genuinely powerful for live transcript streaming without a custom backend.

What's next

Native Zoom/Google Meet/Teams bot integration (auto-join, auto-record), team workspace sharing with role-based access, and a mobile companion app for VAs who take calls on the go. MeetSummary started as my personal solution to admin overload. It has now become a real productivity engine I use every day, and I’m excited to keep improving it for other Virtual Assistants.

Built with determination from Lagos, Nigeria.

Built With

  • clickup
  • cloudflare-r2
  • deepgram-nova-3
  • google-sheets
  • hubspot
  • lucide-react
  • microsoft-teams
  • motion
  • notion
  • openrouter
  • pendo
  • radix-ui-primitives
  • react-18
  • recharts
  • salesforce
  • shadcn/ui
  • slack
  • supabase
  • tailwind-css-v4
  • typescript
Share this project:

Updates