Present is a custom Voice AI application and real-time meeting producer built with 
Next.js. In this talk I’ll cover how I refactored the codebase from the legacy three-agent architecture into a simplified voice + steward system. The production pipeline now runs as two lightweight Node processes: a real-time voice agent using LiveKit’s Agents API that listens to room audio, transcribes speech and calls UI tools, and a conductor that delegates to domain stewards running on the server. Stewards read state from Supabase, reason holistically with the latest LLMs (e.g. Claude Haiku via the CANVAS_STEWARD_MODEL env var), and emit structured UI patches. A browser ToolDispatcher executes these actions in TLDraw and React, enabling collaborative canvas editing. I’ll also show how I integrated Model Context Protocol (MCP) to connect to other AI services, added canvas branding and improved the agent tool catalog, and cleaned up the ‘duct-tape’ code from the summer. Live demos will include voice-driven component creation, flowchart steward updates, and multi-agent cooperation.
