SAIF Meeting Co-pilot Runtime

A local meeting co-pilot that writes live diligence notes into SAIF's CRM during founder calls. It uses a fine-tuned local Qwen 2.5 3B model to generate notes from a running transcript of Zoom audio, and writes those notes into the CRM via Liveblocks as they're produced.

Architecture

Two processes communicating over a local Unix socket:

• Python transcriber — captures system audio via BlackHole, runs MLX Whisper for transcription, runs the fine-tuned model for bullet generation
• Node writer — maintains a Y.Doc synced with Liveblocks, writes bullets as ProseMirror nodes, serves the control panel UI

Both processes are orchestrated by launch.sh. Start with one command, stop with Ctrl+C.

Prerequisites

• Spec A adapter — a trained LoRA adapter must exist at ../saif-copilot-training/adapters/
• Liveblocks secret key — from the SAIF CRM Liveblocks dashboard
• Supabase access — URL and service role key for the SAIF CRM database
• Node.js 20+
• Python 3.12+ (arm64) — MLX requires native Apple Silicon Python. Use Homebrew: brew install python@3.12
• BlackHole 2ch — for capturing Zoom audio (see setup below)

BlackHole setup (for Zoom audio capture)

1. Install BlackHole 2ch:

   brew install blackhole-2ch

   Or download from https://existential.audio/blackhole/

2. Open Audio MIDI Setup on macOS (search Spotlight for it).

3. Click the + button at the bottom left → Create Multi-Output Device.

4. Check both BlackHole 2ch and your normal output (speakers or headphones) in the device list. This lets you hear Zoom audio while BlackHole also captures it.

5. In Zoom → Settings → Audio → Speaker, select the Multi-Output Device.

6. The runtime will automatically detect BlackHole 2ch as its input source. If BlackHole is not installed, it falls back to the default microphone (useful for testing).

Setup

# 1. Install Node dependencies
npm install

# 2. Create arm64 Python virtualenv (MLX requires Apple Silicon native Python)
/opt/homebrew/bin/python3.12 -m venv venv
./venv/bin/pip install -r requirements.txt

# 3. Create .env.local from example and fill in values
cp .env.example .env.local
# Edit .env.local with your keys

Running

./launch.sh

This will:

1. Verify your environment is set up correctly (including arm64 Python check)
2. Start the Node writer (HTTP + WebSocket server)
3. Open the control panel in a floating Electron window
4. Start the Python transcriber (loads Whisper model, begins audio capture)

Whisper model selection

Default: mlx-community/whisper-base-mlx (fast, good for M1 Pro).

If transcription accuracy is poor, switch to the larger model in .env.local:

WHISPER_MODEL=mlx-community/whisper-small-mlx

The model is downloaded on first run from Hugging Face Hub and cached locally.

Current state (Session 3)

Session 3 adds real audio capture and transcription:

• Real audio capture via BlackHole 2ch (or fallback to default mic)
• Real transcription via MLX Whisper — transcript appears live in the control panel
• Verbatim bullets — every ~20 seconds, the latest transcript chunk is written to Liveblocks as a TRANSCRIPT: ... bullet (temporary, replaced with model inference in Session 4)
• Real Liveblocks writes — bullets appear in CRM in real time (from Session 2)
• Status indicators — the UI status dot reflects what's happening (listening, transcribing, idle, error)

What subsequent sessions add

Session	What it adds
4	Model inference with the fine-tuned Qwen adapter — real diligence notes
5	Safety nets, adaptive cadence, dry-run mode

Known limitations

• No real model inference yet — bullets are verbatim transcript, prefixed with TRANSCRIPT:
• No output sanity checking (Session 5)
• No adaptive cadence (Session 5)
• The "Pause generation" toggle has no effect yet (nothing to pause until Session 4)
• Silent audio chunks are skipped — if RMS is below threshold, no transcription runs