System Architecture

Inspiration

There seem to be opportunities on prediction markets to make money because the markets are still young, so instead of manual trading, we wanted to try to script it. Also, we wanted to play around with LLM agents and see how we can empower them to be even more capable.

What it does

This is a system that interacts with Kalshi. There are the normal services like an execution service to communicate with Kalshi, a database service to store historical data and trade states, a risk analysis service to make sure trades don't make us go bankrupt, a backfill service to gather historical data, and a bunch more. However, the part that brings it even higher up is our autonomous QR + QD loop. If a certain signal occurs, an LLM agent is spawned to utilize our tools to analyze our historic data to ensure there is a pattern to be exploited, it then creates a trading service with the algorithm to exploit the pattern, and runs it through a backtester. It will then refine if necessary the strategy and try again unless deemed unworthy of refining. If it passes the backtester, this strategy is deployed and is actively used to trade the Kalshi markets.

How we built it

Oh my gawd. Lots and lots. It will be useful to just look at the github. However, tech stack is c++ and python. Then we used the hackathon sponsors (Tiger, Backboard, Grok, Vultr) to host our prod version of the project.

Challenges we ran into

There are so many services. We couldn't refine them to what we wanted them to be in the timeframe of this weekend.

Accomplishments that we're proud of

The services did come together, even if the system won't be able to output a successful trading algorithm any time soon.

What we learned

In order to build out all these services, we had to cut corners and just trust the LLM understood us and merged the code. However, it seems that slight deviations in interpretations can have a large later on impact and so that's why the current state of the project is not exactly in line with our initial vision.