-
-
AI debate arena
-
Gemini generated analysis based on the AI debate with prediction for outcomes
-
Gemini generated analysis based on the AI debate with prediction for outcomes
-
Search feature
-
Markets view trending
-
Markets view finance
-
Log in
-
User authentication through email code link
-
Saved markets (favorites)
-
Choosing models for AI debate
-
database
Inspiration
Prediction markets are full of takes, but no place to watch top-tier models actually argue their positions. We wanted a way to turn raw market data into a live, intelligent debate where every model exposes its reasoning instead of hiding behind a single percentage.
What it does
Polydebate pulls real Polymarket questions, lets you pick your lineup of LLMs, and makes them fight it out round-by-round with audio and reasoning. At the end, it generates a clean breakdown of who predicted what, why they disagreed, and which arguments actually moved the needle.
How we built it
We wired the Polymarket Gamma API into a Next.js + Flask pipeline, streamed live messages over SSE, and used ElevenLabs to synthesize real-time model “voices.” Gemini stitched AI debates into structured summaries, while our custom orbital UI gave each model a personality without turning the interface into clutter.
Challenges we ran into
Coordinating multiple models speaking, thinking, and generating audio at the same time was chaos at first. Streaming responses, syncing audio, and keeping the UI buttery smooth required a ton of debugging around race conditions and timing mismatches.
Accomplishments that we're proud of
We are team of 2 people and built a full multi-model live debate system in under 24 hours (vibe coding rocks), complete with audio, UI animation, and long-form reasoning analysis. Watching models argue literally speaking over each other felt like the moment we knew this idea actually works.
What we learned
LLMs don’t just disagree they do it with style when given structure and a stage. We also learned a lot about the quirks of live audio generation, SSE streaming stability, and how to design UIs that feel alive without being distracting.
What's next for Polydebate
We’re building a full leaderboard that tracks which AI models actually win debates, predict markets the most accurately, and consistently sway other models during argument rounds. The goal is to turn Polydebate into the first real arena for “AI persuasion performance,” where you can see not just who was right, but who argued their way there better than anyone else.
Built With
- css
- elevenlabs
- flask
- gammaapi
- geminiapi
- git
- javascript
- next.js
- openrouter
- python
- sqlalchemy
- typescript

Log in or sign up for Devpost to join the conversation.