IntelliDebate AI

Two AI models debate any topic—one FOR, one AGAINST—across live rounds. A third AI judge picks the winner and names each side's best point. Choose intensity from logical to brutal.

Comment

Inspiration

I kept noticing how one-sided most AI answers feel — you ask a question and get a single, confident take. Real understanding comes from hearing both sides clash. I wanted to see an argument get stress-tested, not just read a summary, so I built an arena where two AIs actually fight it out.

What it does

IntelliDebate AI takes any topic you enter and runs a live debate between two AI models — one arguing FOR, one AGAINST — across multiple rounds. You pick which side opens, how many rounds, and the intensity (from composed and logical all the way to brutal). A third AI acts as an impartial judge: it reads the full transcript, writes a verdict, declares a winner, and surfaces each side's single strongest point. Every debate is saved to a local history you can revisit.

How we built it

The whole thing is a single self-contained HTML file — no build step, no backend. The frontend is hand-written HTML/CSS/JS with a custom design system (Sora + Space Mono, gradient mesh backgrounds, animated argument cards). The AI runs on Groq's compound models: compound-mini and compound debate each other, and compound-mini judges. An orchestrator function manages turn order, streams responses word-by-word for a live feel, and handles the verdict and takeaways as separate model calls. History persists through localStorage.

Challenges we ran into

  • Rate limits: Groq's free-tier token limits were easy to hit, especially with the compound models pulling in large contexts. I built retry logic that reads the server's "try again in Ns" hint and backs off automatically.
  • Empty responses: Reasoning models sometimes spend their whole token budget thinking and return blank content, so I added fallbacks to read the reasoning field and tuned reasoning_effort to stay fast.
  • Keeping AIs on-side: Models love to hedge and see both perspectives. Tight system prompts force each one to stay firmly committed to its position in 2–3 sharp bullets.
  • Live feel without true streaming: I simulated streaming with a word-by-word reveal so the debate feels alive even on non-streaming calls.

Accomplishments that we're proud of

A complete, polished product in one file that anyone can open and run instantly. The intensity modes make it genuinely fun to watch, the judge verdicts feel fair, and the whole experience — from landing page to verdict to saved history — feels like a real shipped app, not a hackathon prototype.

What we learned

How to orchestrate multiple AI models in a single coherent flow, how to write prompts that hold a model to a stance, and how to handle the messy realities of real APIs — rate limits, empty completions, and reasoning-token budgets. I also got a lot sharper at building a cohesive design system from scratch.

What's next for IntelliDebate AI

  • Real token streaming for instant responses
  • Let users pick which models debate
  • A "join the debate" mode where a human argues against the AI
  • Shareable debate links and exportable transcripts
  • Topic suggestions and a leaderboard of the spiciest debates
Share this project:

Updates

Submission history