Voice Command Dashboard

Bloom.Launch.Strategy.Meeting.Updates.mp4

Real-time meeting intelligence dashboard that listens to conversations and extracts structured artifacts — decisions, action items, risks, dependencies, entities, and more.

Try the demo: Click "Play Demo" to hear a sample meeting and watch the system extract insights in real time.

How It Works

Browser Mic / Demo Audio
        │
        ▼ PCM 16kHz via WebSocket
  ┌─────────────┐
  │ SessionAgent │  (Cloudflare Durable Object)
  │              │
  │  STT ────────┼──► Mistral Voxtral Realtime
  │              │
  │  Fast LLM ───┼──► Per-sentence extraction (low latency)
  │  Deep LLM ───┼──► Periodic consolidation (tool-calling loop)
  │              │
  │  SQLite ─────┼──► Full transcript + artifact persistence
  └──────┬───────┘
         │ WebSocket
         ▼
   React Dashboard

The system extracts items across 5 categories:

Category	Item Types
Actions	action items, commitments, follow-ups
Decisions	decisions, approvals, objections, change requests
Risks & Blockers	risks, blockers, issues, constraints
Planning	dependencies, milestones, deadlines, requirements, metrics
Open Threads	open questions, answered questions, parking lot items

Plus relations (depends_on, blocks, supersedes...), entities (people, teams, systems...), and topics.

Quick Start

Prerequisites

• Node.js 18+
• A Cloudflare account (free tier works)
• API keys for at least one LLM provider + Mistral (for STT)

1. Clone and install

git clone https://github.com/YOUR_USERNAME/voice-command-dashboard.git
cd voice-command-dashboard
npm install

2. Set up API keys

# Login to Cloudflare
npx wrangler login

# Set required secrets
npx wrangler secret put MISTRAL_API_KEY     # Required for speech-to-text

# Set ONE of these depending on your LLM_PROVIDER:
npx wrangler secret put GEMINI_API_KEY      # If using gemini (default)
npx wrangler secret put ANTHROPIC_API_KEY   # If using anthropic
npx wrangler secret put OPENAI_API_KEY      # If using openai

For local development, create a .dev.vars file:

MISTRAL_API_KEY=your-mistral-key
GEMINI_API_KEY=your-gemini-key

3. Configure LLM provider (optional)

Edit wrangler.toml [vars] to change the LLM provider and models:

[vars]
LLM_PROVIDER = "gemini"                          # gemini | anthropic | openai | workers-ai
LLM_FAST_MODEL = "gemini-3-flash-preview"        # Fast per-sentence extraction
LLM_DEEP_MODEL = "gemini-3-flash-preview"        # Deep periodic consolidation
STT_PROVIDER = "mistral"
STT_MODEL = "voxtral-mini-transcribe-realtime-2602"

4. Deploy

npm run deploy

This builds the frontend with Vite and deploys everything (Worker + static assets) to Cloudflare. Your dashboard URL will be printed at the end.

5. Local development

npm run dev            # Wrangler dev server on :8787
npm run dev:dashboard  # Vite dev server on :5173 (proxies API to :8787)

Open http://localhost:5173.

Tech Stack

• Runtime: Cloudflare Workers + Durable Objects
• Agent SDK: agents (WebSocket state sync, SQLite persistence)
• STT: Mistral Voxtral Realtime
• LLM: Gemini, Anthropic, OpenAI, or Workers AI (configurable)
• Frontend: React 19 + Vite 7 + AudioWorklet
• Language: TypeScript (strict)

Deep Analysis

The deep analysis engine uses a tool-calling loop for structured edits:

• apply_operations — upsert/delete items, merge duplicates, manage relations and entities
• finalize — signal completion

This runs periodically over the full transcript and can also be triggered on-demand. Each deep run's changes can be undone.

If a provider doesn't support tool-calling, it falls back to single-shot JSON extraction.

Demo Audio

A pre-recorded 3-minute meeting demo is included (src/dashboard/public/demo-meeting.wav). Click "Play Demo" in the UI to stream it through the full pipeline.

To regenerate the demo audio with different voices:

ELEVENLABS_API_KEY=your-key bash scripts/generate-demo.sh

Scripts

Command	Description
npm run dev	Start Wrangler dev server
npm run dev:dashboard	Start Vite dev server
npm run build	Build frontend + typecheck
npm run typecheck	TypeScript type checking
npm run deploy	Build + deploy to Cloudflare

License

MIT