MeetFlhow

AI meeting intelligence: upload or record a meeting, get a transcript (Deepgram), and a summary, action items, decisions, open questions, and sentiment (Gemini).

Stack

Next.js 14 (App Router) · Neon Postgres · Drizzle ORM · NextAuth.js (Auth.js v5, Google OAuth) · Deepgram + Gemini · Tailwind CSS v4 + shadcn/ui

Setup

1. Install dependencies (already done if you're reading this from the generated project):

   npm install
   

2. Create a Neon database at neon.tech and copy its connection string.

3. Create a Google OAuth client at the Google Cloud Console:

   • Authorized redirect URI: http://localhost:3000/api/auth/callback/google (and your production URL equivalent)

4. Copy .env.example to .env.local and fill in real values:

   cp .env.example .env.local
   

   • DATABASE_URL — Neon connection string
   • NEXTAUTH_SECRET — generate with openssl rand -base64 32
   • NEXTAUTH_URL — http://localhost:3000 locally
   • GOOGLE_CLIENT_ID / GOOGLE_CLIENT_SECRET — from step 3
   • DEEPGRAM_API_KEY — from console.deepgram.com
   • GEMINI_API_KEY — from aistudio.google.com/apikey

5. Push the schema to Neon:

   npm run db:migrate
   

   (A migration is already generated at drizzle/0000_massive_johnny_storm.sql. db:migrate applies it. Use npm run db:generate after future schema changes, and npm run db:studio to browse data.)

6. Run the dev server:

   npm run dev
   

Voice recognition service

Speaker voice recognition (enrolment + identification) runs through a self-hosted Python microservice at voice-service/ using SpeechBrain's ECAPA-TDNN speaker embedding model — no API keys or usage limits. It must be running alongside the Next.js app in dev for voice enrolment/identification to work; without it, those features fail closed (the app still works, meetings just fall back to Gemini's speaker guesses).

Set up once:

cd voice-service
python -m venv .venv
.venv\Scripts\activate   # or `source .venv/bin/activate` on macOS/Linux
pip install -r requirements.txt

Requires the ffmpeg binary on your PATH (used to normalize uploaded audio before embedding).

Then, side by side with npm run dev, run:

npm run voice-service

which starts uvicorn main:app --port 8001 --reload. Set VOICE_SERVICE_URL in .env.local (defaults to http://localhost:8001, see .env.example) so the Next.js app knows where to reach it.

Notes

• The Deepgram and Gemini calls in /api/meetings/upload and /api/meetings/analyze can take a while for long meetings — both routes are configured with maxDuration = 300. On Vercel, this requires a plan that supports extended function durations.
• Session strategy is database-backed (Auth.js v5 + Drizzle adapter), so signing out anywhere invalidates the session everywhere.