Inspiration
Humans discover products, then gossip about them. We simulate that dynamic to improve product discovery.
What it does
Input your product—our agents rank it against competitors.
TinyHumans generates diverse personas based on real population stats. TinyFish fetches up-to-date product info, then agents discuss and vote on what’s best.
You get market insights from simulated “word-of-mouth.”
How we built it
- Qwen-2.5-7B for fast, local agent simulation (no rate limits)
- TinyFish & Exa for Web Search
- FastAPI backend
Challenges
LLM APIs were too slow, expensive, and rate-limited. We switched to a self-hosted Qwen-7B on Modal.
Issue: occasional Chinese outputs.
Accomplishments
- Clustered personas with k-means → agents only “gossip” within their group (realistic social dynamics)
- Orchestrated 1k+ conversations across 200 AI agents
What we learned
How to generate realistic personas from population data.
What’s next
- Scale to thousands of simulated users for stronger signals
- Add real-time feedback loops using actual user data to calibrate agent behavior
- Build a dashboard for companies to run simulations and track product perception over time
Built With
- openai
- python
- qwen
- tinyfish
Log in or sign up for Devpost to join the conversation.