About: skills demonstrated and the builder behind Off The Pace

Off The Pace mapped to concrete data-engineering competencies, each linked to its evidence in the repo: data modelling, data contracts, orchestration, ML engineering, CI/CD, IaC, observability, and cost.

Off The Pace is a portfolio project built to clear a production data-engineering bar end to end, from raw API pulls to a deployed, monitored, browser-native app. This page maps the work to the competencies a reviewer is grading for, with a link to the real artifact behind each one. Nothing here is a keyword without an artifact attached.

Skills demonstrated

Competency	Demonstrated by	Evidence
Data modelling	Medallion warehouse (bronze → silver → gold), dimensional marts, 60 dbt models in one topological pass	Transform layer
Data quality & contracts	443 dbt tests, the CI-enforced additive-identity invariant, schema contracts	The CI Contract
Orchestration	Staged pipeline DAG with retries, source-freshness checks, and build-over-build profiling	Data pipeline
ML engineering	XGBoost trained and exported to ONNX with train/serve parity and calibrated intervals	ML layer
CI/CD	Deploy / promote / rollback, PR preview channels, drift gates as merge blockers	Deployment
Infrastructure as code	Terraform for the GCS bucket, Workload Identity Federation, and a billing budget	Architecture decisions
Observability & cost	Sentry, web-vitals RUM, synthetic monitors, SLOs, and a zero-server cost model	Observability · Performance

For the narrative version of how these fit together, see Engineering highlights (the sixty-second tour) and How this clears a production bar (the deep read).

What this project is

A full-stack F1 analytics platform that answers a question lap time alone cannot: when a car is off the pace, exactly why? It ingests seven seasons of telemetry, models it through a dbt + DuckDB warehouse, trains five ML models, and serves 30 interactive features in the browser with no server, no login, and no per-user serving cost. Every lap is decomposed into seven additive, physically-grounded components, so lost time is attributed to a named cause rather than a feeling.

The engineering thesis is that the interesting numbers should be enforced, not asserted: the seven-term decomposition closes to zero on every lap because a CI test fails the build otherwise, and the model that serves in the browser is provably the model that trained because ONNX parity is tested to atol=1e-5.

About the builder

Built by Justin Clarke. The project doubles as a working demonstration of an end-to-end data and ML platform: data contracts and lineage, reproducible drift gates, ML with calibrated uncertainty and honest out-of-sample validation, CI/CD with rollback, infrastructure as code, and observability with SLOs.

Portfolio site

More work and contact details.

GitHub repository

The full source: ingestion, transform, ML, app, and infra.

Launch the app

All 30 features, running client-side.