Work With Me
I build datasets that do not exist yet, and the analysis that makes them worth having.
What I do
Most analysis restates numbers someone else compiled. The work on this site does the opposite: every figure in the research section comes from a script in a public repository, run against primary sources, with the raw rows published alongside the conclusion. That is the same thing I do on commissioned work.
The datasets behind this site are archived and citable at doi.org/10.5281/zenodo.22654387, released under CC BY 4.0, with the harness that produced them under MIT. You can read the code before deciding whether to trust the numbers — which is the point.
Datasets that do not exist yet
The interesting number is usually the one nobody has compiled. I pull it out of primary documents — SEC filings, government statistical series, published policies — and hand back clean rows with the extraction code and the source quote attached to each figure.
Platform take rates, from eight companies' annual reportsModel evaluation you can reproduce
Benchmarks for language models, run as measurements rather than demos: fixed seeds, explicit context length, repetitions to separate a reliable answer from a lucky one, and the failed runs published alongside the good ones. If a result cannot survive being re-run, it is not a result.
320 generations testing whether naming an account changes the answerCalculators and data visualisation
Interactive tools that make a trade-off legible in dollars, and charts that argue a point instead of decorating one. Every chart ships a data-table fallback, so it works for screen readers and for anything parsing the page.
The tools on this siteAnalysis written by a person
A model can produce a dataset. It cannot notice that a filing's discussion quotes one segment's volume while its revenue is consolidated, which is the kind of error that quietly becomes a published number. I write the analysis, and the method section says exactly how each figure was checked.
How the work is checkedHow a project usually starts
Send the question, not the specification. “What do delivery platforms actually keep?” is a better brief than a schema, because the shape of the dataset should follow from the question rather than the other way round. I will tell you whether the data can be got at all — sometimes it cannot, and finding that out early is worth more than a hopeful estimate.
You get the cleaned dataset, the code that produced it, and a written method covering what was checked and what the numbers cannot support. Scope and timing depend on what the sources allow, so those are worth a short conversation rather than a number on a page.
Get in touch
Email [email protected] with the question you want answered and any deadline you are working to. If it is easier to start from the existing work, the research section is the fastest way to see how a finished piece reads, and the repository shows how it was built.