PocketPlay | Devpost

Inspiration

We chose PocketPlay because we found great potential in the relatively underexplored market of football play simulation. The tool we envisioned fit nicely into the broader idea of high-performance computing due to the natural variance of football play analysis. The possibility of doing something that combined HPC and our passion for football in an analytical sense was something we wanted to work on.

What it does

PocketPlay opens with a welcome page and a built-in Auth0 sign-in process.

The heart of PocketPlay is in the play simulator backed by our machine learning model. After signing in, users are prompted to choose between the coach and NFL modes. The coach mode opens a page with a graphical representation of the football field with offensive and defensive markers representing players. The user can affect the position of the markers by selecting the offensive and defensive schemes from the drop down menus to the right of the football field. This will set up formations for the football players that can then be simulated by the model to identify how often an offensive or defensive scheme is successful. Our tool outputs success rates, yards per attempt and counts of how many plays were either complete, incomplete, turnovers, or touchdowns. There is also an AI agent that can evaluate the play for users and provide additional feedback on what play or feature may work better against a specific type of defensive scheme. The graphical representation of the field also includes an animation that depicts the movement of wide receivers running their specified routes. Furthermore, there is a POV feature that operates as a “street view” depicting the result of the play through a 3D graphical representation as an additional tool for coaches to utilize. Each player’s profile can also be adjusted based on their strengths and weaknesses leading to both individual and team ratings that affect their ability to execute on either offense or defense.

The NFL mode of PocketPlay functions similarly to the coach mode but instead it acts as a tool for fans, analysts, teams and anyone interested to run plays with NFL teams. This mode has drop down menus to choose teams and the offensive and defensive markers have the profiles and names of existing NFL players based on a Madden 26 dataset. This mode also has a machine learning play simulator that outputs valuable statistical information like success rate and yards per attempt. The mode also simulates the outcome of thousands of iterations between teams which can be used to estimate win probability and identify strengths and weaknesses of NFL teams. This has the potential to be especially useful for the NFL postseason which can inform the general public on potential outcomes between matchups.

How we built it

For this project we used Next.js and react as frameworks for the logic part. Then we used Auth0 for the sign in feature. We used the Google API key for our agent. We utilized cursor AI to build our agentic AI tool.

Challenges we ran into

In the beginning we wanted to incorporate a larger amount of AI, so it would've enabled the user to speak directly with the chatbot and have it run all the simulations based on user preference. Instead we ended up going with a drop down menu with pre-stated plays, which forced us to limit the amount of plays we could include in our website. Additionally, for our machine learning play simulator, it was tricky to find the appropriate datasets to train our model. In our perfect world, we would have liked to have access to play-by-play datasets with details of every NFL offensive and defensive formation but data of this quality is often only available to professional teams or is behind a heavy paywall. We had to settle for less detailed but sufficient information from nflfastR for the current version of the tool. Additionally, we had some UI issues with the players not lining up properly on the field which required a lot of debugging.

Accomplishments that we're proud of

We are proud of training the machine learning model to create accurate analyses of game plays even though a lot of data was not readily available to us. This caused issues since a lot of our website relied on this feature. We had to think on our feet and come up with a way that we could get a reliable data set and represent what we wanted. Additionally, we utilized existing Madden statistics as valuable strength evaluation data and this proved to be successful since this Madden data also comes from actual sports data and has already been combined to serve our purpose. This added a lot to our NFL mode which is geared towards analysts, fans, and the general public to provide well organized and visually pleasing data.

What we learned

This project taught us a lot about data analysis, machine learning, and creating passion-driven products that can appeal to anybody. We learned a lot about UI development and collaborating on front end stuff while working on Git. Like we said before we also learned a lot about machine learning and how to properly train the model. And overall we all gained a new understanding of what HPC really is and how it can be utilized in a variety of use cases.

What's next for PocketPlay

In the future we definitely want to be able to integrate data with a higher accuracy ceiling and detailed play-by-play information so that our product has an even more competitive chance with our audiences. We would also like to learn more about 3D rendering and animation to keep developing our 3D view mode. Another thing we might want to circle back to integrating more AI like we wanted to do initially, to make the website even easier for the user to use. With time, our playbook with offensive and defensive schemes can be expanded further to fully encompass the broad landscape of football decision-making. This is a process that begins with more data that can better train our models.