Spoorters

AI agents that outthink the odds.

Comment

Inspiration

We saw an opportunity in fanduel and polymarket specificially with sports betting. AI Agents are on the rise yet big quantitive firms like SIG, Jane Street, etc are all hesitating incorporating AI agents in their market predictions.

What it does

In short our app transforms any NBA prediction question into a network of AI agents that analyze real data through MCP servers and outputs a simple YES/NO output along with probabilities. With the probabilities, a trade will then be placed.

How we built it

We created an MCP and hosted it with Dedalus-Labs and FastMCP to get it production ready fast as possible that has information on historical NBA games and player data. We then built a frontend that directly uses the Kalshi API to get live bets that are live on the market so traders can directly choose a YES/NO question to bet on. That question is then handled through our framework which we implemented to break it down into 5 smart sub-questions and then each of those into 5 sub-questions so in total 131 questions. We then send out AI Agents equipped with the NBA MCP along with a deep research MCP to answer each subquestions so a more accurate model of the true probability of the main question can be reconstructed which we did through Bayes Theorem.

Challenges we ran into

Some Challenges we ran into were connecting our website with our MCP server as well as building the machine learning Question decomposition framework which we had to put on hold due to long training times. We couldn't implement direct trades due to us not having funds.

Accomplishments that we're proud of

Accomplishments we are proud of is getting the MCP servers to work, creating a prototype of the Question decomposition framework in the time frame we had.

What we learned

We learned to manage that we need to get to work faster instead of waiting until 9pm at night to start. We also learned that for AI agents a large amount of data is needed with an access to an enormous amount of API for this product to scale.

What's next for Spoorters

Spoorters will begin to train the unsupervised model we have to decompose questions into smaller and smallers questions through Pytorch to improve our question decomposition system. We would also improve our site and functionality and add direct trades.

Built With

Share this project:

Updates