Vibeathon

• 

  Our virtual hackathon site

• 

  Vibeathon Arena for all our projects by ai

• 

  Timeline for "Vibeathon HackIllinois"

Inspiration

It all started with one dumb-fun question: what if you actually locked 50 of the most brilliant, chaotic, and legendary minds in history inside the same hackathon?

Picture Elon roasting Napoleon’s product roadmap, or Kanye dropping a wild creative bomb on Gandhi. We were obsessed with Stanford’s Generative Agents paper and the raw energy of real hackathons, so we decided to build the whole thing as a multi-agent social simulation. Every persona has its own personality baked in — not just a name, but actual psychological traits.

What We Learned

• Prompt engineering 50 unique humans is no joke. We went deep: childhood stories, biggest controversies, signature quotes, risk tolerance, empathy levels, stubbornness scores… the works. A persona isn’t just “act like Elon,” it’s a whole cognitive model.
• Getting agents to actually collaborate instead of talking past each other was brutal. We had to build a strict session flow: Pitch → Debate → Decision, and force every agent to throw at least one real objection.
• ELO voting is shockingly elegant. We straight-up stole the chess formula and it just works:

$$E_A = \frac{1}{1 + 10^{(R_B - R_A)/400}}$$

$$R'_A = R_A + K \cdot (S_A - E_A)$$

How We Built It

Three main pieces:

1. Social Simulation Engine (social-simulation)
   Python async beast that runs 7 full rounds. Agents go through free chat → team formation → brainstorming → polishing → final pitch. Persistent memory across rounds via Supermemory.ai.

2. LLM Inference Backend (modal-backend)
   Deployed on Modal with vLLM serving Qwen on 4× B200 GPUs (tensor parallel). Full OpenAI-compatible API, streaming, handles 55 concurrent agent chats without breaking a sweat.

3. Arena Frontend (arena-web)
   Next.js 16 + React 19 on Cloudflare Workers. 1v1 voting, live ELO updates, Neon Postgres in the back.

Challenges We Faced

• Keeping personality consistent across hours of conversation without turning into robots. Fixed it with locked persona params + a context manager that protects the core traits but lets ideas evolve.
• Temperature was everything. Too high early = total chaos, too low late = boring. So we ramped it down from 0.95 in round 1 to 0.60 by round 7.
• Running 50 agents in parallel without melting the GPU budget. Lots of async retries, careful orchestration, and some “please don’t die” logic.

Built With

Built With

• cloudflare
• modal
• neon
• nextjs
• python
• supermemory
• tailwind
• typescript
• vllm
• wrangler