notulen.online

The client is an Indonesian consulting firm with 20+ client meetings a week. Associates spent an average of 4 hours a week just transcribing and cleaning up minutes. Existing tools (Otter.ai, Fireflies, Tactiq) are English-first, billed per seat at a high price, and their bots often get kicked out of meetings because the name looks unfamiliar. They needed a local platform: Indonesian-first, sensibly priced, and self-hostable for sensitive clients.

Four overlapping problems: (a) a bot that can join Meet/Zoom/Teams without being spammy — handling permissions, the recording indicator, and exiting gracefully when the host ends the meeting; (b) Indonesian transcription across Javanese/Sundanese/Betawi accents — default Whisper accuracy drops; (c) streaming real-time transcripts to multiple viewers with < 2 seconds of delay; (d) AI output credible enough for formal minutes — a different tone from a casual summary, with decisions, owners, and deadlines.

Next.js 16 + Laravel 12 (Horizon queues, Reverb WebSocket broadcasting) + FastAPI Python for Whisper/Moonshine/pyannote + a separate Node Puppeteer bot service. Audio in MinIO (S3-compatible), transcripts + embeddings in MySQL 9, vector search for the knowledge base. Bots scale horizontally: 10 concurrent per worker, the 11th gets a graceful 503 + auto-queue. A ready-made Flutter mobile app for field teams recording voice memos.

MVP live in 12 weeks. Smoke test: 1 hour of audio → transcript + minutes + action items ready in < 2 minutes end to end. Pilot tenants (the consulting firm + 2 Jakarta startups) are onboarded for internal meetings — saving roughly 3 hours per associate per week. Next: native Slack/Notion integration to auto-push minutes to each workspace, plus further Indonesian fine-tuning of Whisper.