> ## Documentation Index
> Fetch the complete documentation index at: https://offthepace.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.
# Off The Pace: F1 Lap Decomposition Analytics
> Every F1 lap split into seven physically-grounded components, served by a fully client-side pipeline: FastF1 ingestion, dbt + DuckDB transforms, XGBoost models exported to ONNX, and a React app that runs entirely in the browser.
Off The Pace answers the question lap time alone can't: **when a car is off the pace, exactly why?** Every lap is split into seven named, additive components fuel, tyre degradation, track rubber, ambient conditions, constructor pace, dirty air, and driver skill so you can see precisely how much of the gap is the car, the tyres, the conditions, or the driver. Underneath that product is a full data and ML platform: a CI-enforced decomposition, a 60-model dbt warehouse, ONNX models with train/serve parity, and a zero-server browser app.
## Are you here to…
The sixty-second tour for a reviewer: the eight things to notice, then the skills demonstrated and who built it.
The architecture, the build-time vs request-time split, and how it clears a production data-engineering bar.
Clone the repo and ingest your first race, or launch the live app.
## How the pipeline fits together
Raw timing data becomes an interactive chart through four layers, each documented in its own tab:
```mermaid theme={null}
flowchart LR
A["Sources
FastF1 · Jolpica"] --> B["Data
Bronze ingestion"]
B --> C["Transform
dbt + DuckDB"]
C --> D["Machine Learning
XGBoost → ONNX"]
D --> E["App
React, in-browser"]
classDef layer fill:#111827,stroke:#e40404,color:#fff;
class A,B,C,D,E layer;
```
Where the timing data comes from and how to ingest it yourself.
The seven-term decomposition: dbt models enforced by a CI contract.
XGBoost models predicting tyre cliff onset and degradation, exported to ONNX.
The browser app every chart and simulator above ships from.
## By the numbers
2018–2024 of F1 timing data ingested end to end.
Every lap decomposed and verified against the additive identity.
620 tests enforce the pipeline on every build.
XGBoost trained, exported to ONNX, running live in the browser.
Interactive visualizations, zero server, sub-10ms queries.
DuckDB-Wasm + ONNX Runtime Web run the whole stack client-side.
## The stack
### Data Pipeline & ML
Telemetry & lap timing API
Championship standings & pit API
Ingestion & orchestration
In-process data warehouse
Transform layer & data tests
Gradient-boosted models
### App & Inference
App framework & build tool
In-browser SQL analytics
In-browser ML inference
### Infrastructure
Gold-mart CDN
Static hosting
CI: tests, drift gates, deploy
## What you can do
Dial in a stint compound, fuel load, and dirty air and watch the trained models project pace loss and cliff risk live.
Every driver re-ranked in equal machinery a counterfactual championship built from the structural pace model.
Any lap, any driver: seven causes stacked into one bar that always closes to zero.
How long each compound lasts before the cliff, by circuit and season.
"Did Mercedes' pit call or Hamilton's driving win São Paulo 2021?"
`fct_lap_residuals` splits every lap of the race into fuel, tyre, rubber, ambient, constructor, dirty-air, and driver-skill components see [the case study](/findings/sao-paulo-2021).
The same physics terms feed the ONNX degradation models that power the live simulator.
The React app queries the gold Parquet mart with DuckDB-Wasm and renders the answer client-side, in milliseconds.
The vocabulary used throughout the docs pace delta, clean laps, the tyre cliff.
Clone the repo and ingest your first race in five minutes.
Real race decompositions, starting with Hamilton vs Verstappen at São Paulo 2021.