WhaleSwarm

Inspiration

We built our repo from scratch of course, but a big inspiration came from another GitHub repo https://github.com/666ghj/MiroFish a recent breakthrough in simulation technology. We thought this was an incredibly cool tool (as did everyone else—it has 45k stars) that could entirely change how we think about LLMs, and we were inspired by the Polymarket track to apply these ideas to prediction markets to harness wisdom of crowds without needing real crowds.

What it does

It simulates hundreds of agents (using a cheap LLM, we use Gemini 2.5 Flash Lite) who interact through a mock Reddit & a mock Twitter, and trade through a mock Polymarket (although the market is always a real, live market that we seed with the up-to-date price using the Polymarket API). They go through several turns where they digest information from other's post, and are occasionally interrupted by the breaking news of a hypothetical you propose (to see how they will react). It simulates the agents by assigning some random characteristics to certain relevant entity types. The relevant entity types are determined from the knowledge graph constructed by an LLM using the documents the web search LLM finds online about your specific market. Thus, each agent comes into the simulation with genuinely unique perspectives (albeit identical training data since we only used one model for simplicity), which allows the group to represent most/all perspectives and converge to a well-rounded estimate of the probability of an event.

Challenges we ran into

This was just an immense project so even though I think we accomplished an amazing amount, there's always more to be done. I'm definitely going to work on this project after YHack is over. I would say my main challenge right now is sleepiness from working for 22 hours straight.

Accomplishments that we're proud of

Novel contribution to the rapidly developing space of swarm intelligence, pulling all of our main ideas together in 24 hours, intuitive UI, effectiveness of simulation.

What we learned

It's really important to have realistic seed personalities if you want to have realistic results from a simulation of many people.

What's next for WhaleSwarm

See if we can make any money applying it to Polymarket sports prediction.

Built With

• and-report-generation.-we-also-used-polymarket?s-clob-and-gamma-apis-for-live-market-data
• and-sql.-we-used-flask-for-the-backend-and-vue-3-+-vite-for-the-frontend
• d3.js
• embeddings
• entity-extraction
• flask
• gemini
• google-cloud-vertex-ai
• javascript
• neo4j
• openai-compatible-apis
• pdfminer
• polymarket-clob-api
• polymarket-gamma-api
• pypdf
• python
• sqlite
• vite
• vue-3
• we-integrated-google-gemini-/-vertex-ai-and-openai-compatible-apis-for-ontology-generation
• with-d3.js-for-graph-and-market-data-visualizations.-the-app-uses-neo4j-for-knowledge-graph-storage-and-sqlite-for-simulation-state.-for-ai-workflows