Pitwall
Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

PITWALL is a real-time race strategy terminal for a problem where the decision can't wait: a strategist has seconds to choose whether to pit, the rival's plan is hidden, and each option can only be judged by what might happen later. PITWALL follows one Formula 1 car against one rival using public timing data. On every lap it predicts when the rival will stop, simulates each of our options to the flag (1,000 races per option, in about 5 milliseconds), and gives the race engineer one stable call and the words to say to the driver. It also shows each option's win chance and where the car would rejoin in the field. The engineer can then ask Meta's Muse Spark about the call in plain English and get a short answer built only from PITWALL's numbers. We tested it on real races. The models were fitted on 2023-24 and tested on 2025 and 2026, using data from 89 races and 99,553 car-laps. Each test race was replayed through the production engine one event at a time, and automated tests confirm that changing future laps can't change a decision. PITWALL's win chance beat the gap-only baseline: a Brier score of 0.111 against 0.129 in 2025, and 0.129 against 0.170 in 2026, where lower is better. We also publish the limits. In the last 10 laps, the gap alone predicts better. Most rival stops come without warning, with 35% to 43% flagged in advance. Each race only shows what the team really chose, so we can't prove that a call would have gained a place. PITWALL is decision support, and it uses OpenF1 data, which is licensed for non-commercial use.