Protagoras v1.5.0: Multi‑LLM Debate + Similarity Matrix
2026-01-15 · Case study
What it is
Protagoras is a multi‑participant debate orchestrator. You configure participants (role + personality + model/provider), choose debate rounds, then run a “meeting” where each participant contributes per round and a moderator synthesizes the final answer. The system also computes an optics-first similarity matrix from the last round of participant responses.
Why it matters
- Multi‑model perspective: Mix providers to get independent reasoning styles and failure modes.
- Meeting feel: Roles + personalities simulate a real decision room (CFO, Legal, Engineer, etc.).
- Metrics optics: Similarity matrix makes convergence/divergence visible at a glance.
How it works (high level)
- Frontend keeps a transcript and orchestrates per-participant calls per round.
- Backend routes calls to the selected provider (OpenAI/Gemini/Anthropic/xAI).
- Final synthesis is produced by the moderator (OpenAI GPT‑4o mini by default).
- Similarity uses a lightweight Jaccard token overlap metric (optics-first; upgradeable later).
What I’d build next
- Export: Download transcript + metrics as HTML/PDF.
- Hard cost caps: Add server-side rate limiting and output token caps per provider.
- Per-round metrics: Show one matrix per round to visualize convergence over time.
- Better similarity: Embedding cosine similarity (still displayed as a matrix).
Demo link: Add your Protagoras Netlify URL here