Agent Arbiter

A live streaming audio system that transcribes speech to text and uses LLM analysis to perform automated tasks.

Architecture

This monorepo contains three packages:

• @arbiter/logging — Pino-based logging utilities
• @arbiter/core — Core services for audio processing, STT, NLP, planning, and tool execution
• @arbiter/workflows — Orchestrator that starts and manages all services

Data Flow

1. Audio Input → WebSocket server receives audio chunks
2. Ingestion → Audio chunks are stored in Redis streams
3. STT Processing → Audio is transcribed to text
4. NLP Analysis → Text is analyzed for intents using LLM
5. Planning → Intents are converted to executable actions
6. Tool Execution → Actions are executed (Jira tickets, action items, etc.)

Quick Start

Using Docker Compose

1. Set environment variables:

   export OPENAI_API_KEY="your-openai-key"
   export JIRA_BASE_URL="https://your-domain.atlassian.net"
   export JIRA_EMAIL="your-email@domain.com"
   export JIRA_API_TOKEN="your-jira-token"
   export JIRA_PROJECT_KEY="PROJ"

2. Start the system:

   docker-compose up --build

The WebSocket server will be available at ws://localhost:3000/ws/audio

Local Development

1. Install dependencies:

   yarn install

2. Build packages:

   yarn build

3. Start Redis:

   docker run -d -p 6379:6379 redis:7-alpine

4. Start the system:

   yarn start:workflows

Configuration

Environment variables:

• REDIS_URL — Redis connection URL (default: redis://localhost:6379)
• OPENAI_API_KEY — OpenAI API key for STT and LLM services
• JIRA_BASE_URL — Jira instance URL
• JIRA_EMAIL — Jira user email
• JIRA_API_TOKEN — Jira API token
• JIRA_PROJECT_KEY — Jira project key

API

WebSocket Audio Input

Connect to ws://localhost:3000/ws/audio and send audio chunks:

{
  "seq": 1,
  "id": "chunk-1",
  "payload": "base64-encoded-audio-data"
}

Health Check

GET http://localhost:3000/health

Development

Project Structure

packages/
├── core/
│   ├── src/
│   │   ├── services/          # Service implementations
│   │   ├── redis/            # Redis adapter
│   │   ├── noop-stt.ts       # STT placeholder
│   │   ├── noop-summarizer.ts # LLM placeholder
│   │   └── ...
│   └── package.json
├── logging/
│   └── src/                  # Logging utilities
└── workflows/
    └── src/                  # Service orchestration

Adding New Tools

1. Implement IToolCommand interface in packages/core/src/
2. Register the command in packages/workflows/src/index.ts
3. Add required environment variables to docker-compose.yaml

Replacing Noop Implementations

• Replace NoopStt with real STT service (OpenAI Whisper, etc.)
• Replace NoopSummarizer with real LLM service (OpenAI GPT, etc.)