Inspired by PewDiePie's experiments running multiple local AI models as a "council" that vote on decisions [1], I wanted to see if you could get similar multi-perspective analysis without a $20k GPU rig.

The approach: use customised system prompts to create distinct personas (methodical professor, creative brainstormer, pragmatic engineer), have them analyse problems independently via parallel API calls, then vote on the best solution using structured outputs (Zod schemas).

Key technical bits: - Structured responses ensure consistent, parseable JSON from the LLM - SSE streaming for real-time UI updates - Parallel processing so personas don't influence each other - Chair Duck orchestrates and breaks ties

Built with Next.js 16, Vercel AI SDK, and Google Vertex AI (Gemini 2.0 Flash).

Live demo: https://rubber-duck-committee.vercel.app/ Source: https://github.com/r-leyshon/rubber-duck-committee Blog writeup: https://thedatasavvycorner.com/notepad/05-rubber-duck-commit...

[1] https://www.pcgamer.com/software/ai/pewdiepie-creates-an-ai-...