03 · Real-Time ML
0.996
R² accuracy
Real-time F1 pit-wall strategy simulator that runs 10,000 Monte Carlo race simulations in under 0.2 seconds.
The Exhibit
01·Problem
Race strategists need to predict how a driver's lap times will evolve as tires degrade and fuel burns off, then simulate thousands of "what if" scenarios — different pit windows, undercuts, safety cars — fast enough to matter in a live race.
02·Approach
- 01
Two-stage ML pipeline: Ridge regression with degree-2 polynomial features trained on 10 circuits predicts a baseline lap time per track (extrapolating to unseen circuits like Monza).
- 02
XGBoost predicts the per-lap residual using race-state features (tyre age, compound, position, fuel load, stint, weather) anchored to the GP median pace.
- 03
Vectorized Monte Carlo engine in NumPy runs 10,000 simulations injecting model residuals and Gaussian noise.
- 04
WebSocket-driven live race state: chaos events (rain, safety cars, crashes) triggered from an iPad "steward" view ripple into new simulations instantly across all connected Pit Wall clients.
- 05
Gemini generates a context-aware radio call narrating outcomes — purely for presentation, no computation.
- 06
FastF1 pipeline extracts historical timing, weather, and telemetry into static JSON for the Next.js frontend, avoiding heavy backend REST payloads.
03·Architecture
Swipe to explore
04·Results
0.996
R² (5-fold CV lap KFold)
0.35s
mean absolute error
<0.2s
for 10K MC simulations
05·Stack
Next.js · FastAPI · WebSockets · Ridge Regression · XGBoost · NumPy · Gemini