This project explores a bit of friendly competition between different models in managing Fantasy Premier League teams. Each model must make team selection, transfer and captaincy decisions throughout the season. The live dashboard tracks their progress, strategies and outcomes in real-time.
This project uses a custom RAG pipeline, with each model receiving a comprehensive prompt with player information, gameweek fixtures, the current team and chip and transfer information. Alongside the rules of FPL, which are quite complex. The prompts are constructed in the same way for each model, with the only differences being the model's current team, chip and transfer information.
Player information is passed through another model (Gemini 2.5 Flash-Lite) first, for augmentation, to determine whether each player is recommended (must have, recommended, rotation risk, avoid) and whether the player is likely to play (fit, minor doubt, major doubt, out). This query searches the web along with seeing the player's stats. This is repeated for every single team (it's quite a lot - 40 queries per gameweek). This information is used to filter down the player list given to each model, to prevent overwhelming them (and to also give a bit of extra context).
This dashboard contains the performance of each model so far and is updated weekly. I also include each team selected with any reasoning. There are also some lab notes, which detail any caveats or interesting observations from each gameweek.
I've written a few detailed posts on this covering some of the technical aspects. Overall though it works! And given the complexity of FPL rules, it's amazing these models actually produce valid teams, let alone ones that are currently better than the average player.