Inspiration
Formula 1 races are won and lost on the pit wall. Race engineers are inundated with thousands of data points from high-performance computing simulations every second, and they have to sift through this data and make critical strategic decisions under immense time pressure. We were inspired by the challenge to bridge the gap between this overwhelming firehose of data and the clear, concise commands a team needs to win. We wanted to build an AI co-pilot that provides answers instead of just showing data.
What it does
Pitwall AI is a real-time race strategy dashboard that transforms complex simulation data into actionable intelligence. It ingests a live feed of potential race outcomes and uses reinforcement learning to analyze the strategic landscape. Instead of forcing an engineer to read dense tables, our dashboard clearly displays what direction the car should move. It's a decision-support system designed to reduce cognitive load and empower race engineers to make faster, more confident calls.
How we built it
We built Pitwall AI as a full-stack, real-time application.
Frontend: The dashboard is a modern React application built with Vite. We used Material UI for the core layout and card components, and Tremor for the specialized data visualization elements like progress bars and metrics.
Backend: A Python backend using the FastAPI framework processes the incoming data stream and communicates with the AI model.
AI Core: We employed a Deep Q-Learning model on a track. At each moment, we use a set of telemetry data on our current vehicle to compute (using HPC) the optimal path that the model should take from the current position of the race.
Simulation: To create a realistic environment, we built a custom F1 simulation in Python using Gymnasium, which generates the live telemetry and race data that our application consumes.
Challenges we ran into
Our biggest challenge was having multiple Deep Q-Learning agents on the same track. We also faced challenges in real-time state management within React, ensuring our dashboard could handle a high-frequency data stream without performance degradation. Finally, creating a non-overlapping, geometrically correct F1 track in our custom Gymnasium environment proved to be a surprisingly complex geometry problem that required several different approaches to solve.
Accomplishments that we're proud of
We are incredibly proud of successfully using neural networks for a high-stakes task like real-time race strategy. We are also proud of the final UI/UX of the dashboard; we believe it effectively presents critical information in a clean, intuitive, and modern interface that genuinely reduces cognitive load. Building a complete, end-to-end system from data simulation to a polished frontend in a single weekend is an accomplishment in itself.
What we learned
This weekend was a deep dive into the practical application of modern AI with High Performance Computing. Without High Performance Computing, our model would not be able to compute the optimal path for the car in time for it to matter to real F1 drivers.
What's next for Pitwall
Pitwall AI has a clear roadmap for the future. The next logical step is to move beyond simulated data and integrate with real-time telemetry APIs from racing platforms. We could also train a custom model on the many hundreds of variables that affect an F1 race and train the model to identify which weights matter the most in deciding the optimal strategy. Finally, we could train a custom model on historical race data to further refine and enhance the strategic recommendations provided by the AI.
Log in or sign up for Devpost to join the conversation.