Writing
20 published pieces, back to January 2025. Most began as something that had to be worked out for a project and was written down so it would not have to be worked out twice. The recurring theme, if there is one: the arithmetic is rarely what goes wrong — the assumptions underneath it are.
Data, Money & Life
Python, data engineering and money. Working code, real numbers, and the occasional post about what did not work.
20 pieces · published since January 2025
Mar 2025
The method behind the collection-rounds work, written out in full: why a greedy route is almost never the good one, and how letting the solver accept a worse answer early is what gets it to a better one later. The animation at the top of this site is this algorithm, running.
Written up from the collection-rounds case study
Read itJul 2025
A national open dataset turned into something queryable, scheduled and actually usable. The same approach underneath the mobility site — public data is only free if you are willing to do the engineering.
Written up from the mobility product
Read itAug 2025
Where a rounded intermediate quietly becomes a wrong answer, traced from plain arithmetic through to model output. The closest thing here to a manifesto: most bad numbers are not wrong calculations, they are careless ones.
Read itLess about what the models can do, more about what they cost and where they quietly fail.
4 pieces
A weekend plugin, eight controlled runs, and a breakeven number that made me use the thing far less than I had planned. Publishing the disappointing result is the point.
Which led to undocumented API headers and a tool to fix what they exposed.
A working example rather than a survey: PyMuPDF for the structure, a model for the parts that resist structure.
Automation that writes in public needs a record of what it said. Includes the BigQuery logging most versions of this leave out.
Small things that remove a recurring irritation, written short on purpose.
3 pieces
python script.py breaks package imports. python -m fixes them. Two minutes, and it removes a class of error people work around for years.
Built because the obvious library does not cover crypto and every workaround was worse than writing one.
The two dialects side by side, for the moment you are certain a function exists and cannot recall which name it has here.
The recurring theme across everything else: the arithmetic is rarely the problem, the assumptions are.
4 pieces
The part of the job that does not appear in the tutorial: cleaning, deciding what a blank means, and discovering the label is wrong.
What fifty attempts at explaining this work taught me about the work itself.
Clustering rows and columns at once, for the case where a group only exists across some of the features and ordinary clustering walks straight past it.
Information gain, explained through a plastic gadget from 1999 that genuinely felt like magic.
Mostly about where the data comes from, which decides everything downstream.
6 pieces
What to use once the free option starts failing on you mid-project.
Coverage, rate limits and what each one actually charges for.
The failure modes that only show up once a project depends on it.
Regime switching as a way of thinking about markets, and where the analogy stops being useful.
A worked classification problem where the cost of a false positive and a false negative are nowhere near equal.
An attempt to understand a genuinely confusing bit of Dutch tax treatment, with the arithmetic shown.
Every tool here runs in your browser and uploads nothing, so you can see what we would see before telling us anything at all. When it is worth a conversation, send the real extract and the question you actually need answered.