Projects

Projects

Systems I built and run myself

Client work is confidential, so it can only ever be described in outline. These are the systems I own outright — which means I can show you the whole thing: the data model, where the language model sits, the running cost, and the part that was quietly wrong for five months before I caught it.

In production · running daily

Velocity AI — analytics on a proprietary binary format

The problem. Endurance training data is locked in FIT files: a binary format emitted by sports watches, one file per session, holding a second-by-second record stream. Every platform that reads it shows the same handful of summary figures. The measurements a coach actually decides from — whether efficiency is improving, whether a steady effort fell apart in its second half, whether an athlete is absorbing the training load or drowning in it — are in none of them, because they are derived from the full stream and not from the summary.

What I built. An ingestion pipeline that parses FIT into a queryable store, a derived metric layer on top of it, and a language model that writes the weekly review a coach would otherwise spend an evening assembling. The draft never reaches an athlete unreviewed — a certified coach edits it first. The system does the reading; the human keeps the judgement.

935sessions parsed, 2022 to date
4 yrscontinuous history, no gaps
8derived metric views
Dailyunattended load and analysis
FIT binary parsing BigQuery Python LLM narrative generation Human-in-the-loop review Partner API integration

The data layer

One row per session, with the full FIT content kept as nested structures — record stream, session summary, laps, events, device info — rather than flattened on the way in. Flattening early is the standard mistake with this format: you cannot ask a question later that you did not anticipate at load time. Keeping the stream intact means a new metric is a view, not a re-ingestion of four years of files. Loads are idempotent and keyed on the source file, so re-running the loader over old data costs nothing.

The metrics the source platform does not give you

  • Cardiac drift — heart-rate-to-pace decoupling across the second half of a steady effort. The most useful fatigue signal in the data, and nothing surfaces it.
  • Efficiency factor — pace per heartbeat, judged against a rolling personal baseline rather than a population norm.
  • Recovery heart-rate drop — how fast the heart comes down once the effort ends.
  • Power-duration curve — the best sustained effort at every duration, built from the record stream.
  • Time-in-zone discipline — how much of a session was spent in the zone it was prescribed for, rather than drifting into the comfortable middle.

Where the language model sits, and where it does not

The model does not touch the arithmetic. Every figure is computed in SQL, then handed over as a table alongside the athlete’s own baselines, and the model’s only job is to turn that into prose someone will actually read — with explicit rules for what counts as a regression and a standing prohibition on congratulating anyone whose numbers did not improve. Keeping computation out of the model is what makes the output reproducible: the same session yields the same numbers every time, whatever the model says about them.

What reaches the API is derived figures and a date. No name, no email address, no GPS trace. At published list prices the generation works out at roughly half a US cent per session on a small model, rising to a few cents on a frontier one — arithmetic any reader can redo from the provider’s own price page.

The part that was wrong for five months

The race-prediction model rested on a power-and-efficiency calculation with a unit mismatch in it, amplified by outliers in the power data. It had been quietly wrong since February. I caught it when it predicted a half-marathon time I had no business running, mid-comeback from injury — and believed it for long enough to be embarrassing.

The replacement is deliberately duller: a Riegel projection drawn only from genuine quality efforts — long enough, hard enough, recent enough — reduced to the median of the best three, paired with a readiness flag comparing the longest recent run against the race distance. It is less clever and it is right. The lesson travels well past running: a model that produces a plausible number from bad units does not announce itself, and the only defence is a sanity check a human would notice failing.

Operating it

The pipeline runs unattended once a day — ingest what is new, compute, generate for any session not yet covered, log it. It has also survived its provider retiring the model versions it was pinned to, which is its own lesson: pin a version and you will be rewritten on somebody else’s schedule, so resolve the model at call time and keep a fallback. The last manual step — exporting session files by hand — was removed by integrating the platform’s own API, so new sessions now arrive on their own.

Why a running project sits on a consultancy site

Because the shape of the problem is the one most businesses actually have.

Operational data that matters is usually sitting in a format nobody has a connector for, holding far more than the dashboard built on top of it admits. The work is the same every time: get it into a store that can be queried, derive the measures the business actually decides on, and only then put a language model in front of it — as the layer that explains the numbers, never the layer that computes them. Velocity AI is that pattern end to end, and it is the one build I can show you in full.

Have data stuck in a format nobody reads?

That is the usual starting point. Tell me what the data is and what decision it should be driving, and I will tell you whether it is a week of work or a quarter.

Start a conversation