Deepgram Nova-3 streams transcription in real-time via WebSocket (sub-300ms, words appear as you speak) If you're offline (because conference WiFi is always terrible), it falls back to local Whisper running in a Web Worker via WASM so it never freezes the UI Rule-based cleanup runs instantly (<50ms) to strip filler words and extract emails/phones from spoken text ("john at gmail dot com" → john@gmail.com) An optional cloud tier does LLM-powered cleanup if connectivity allows (3-second hard timeout, so it never blocks you)
The key insight: at events, you have maybe 30 seconds between conversations. You need something faster than opening the Contacts app and typing. A 10-second voice note that auto-parses into a contact card is that thing.
Stack: React + Vite + Supabase + Deepgram Nova-3 + Whisper WASM. It's a PWA with offline support, so it works even in airplane mode (Whisper handles transcription locally). I've been using it at events for a few months and it's changed how I network I actually follow up now because I have the context of why I wanted to connect, not just a name.
Would love feedback from anyone who goes to conferences regularly. What's your current system for capturing contacts?