Insights

Things worth knowing before you build.

Notes from the work in the portfolio: what public datasets hide, what a modelling shortcut actually costs, and how to tell the difference between a number and a finding. Everything here is measured on something you can open.

Regulatory reporting · 10 September 2026

The average bank is not the sector

Three honest numbers for the capital strength of EU banks, from the same supervisory returns: 16.2%, 18.0% and 22.7%. Only one of them is the sector, and the largest is the one you get by default — averaging a column of ratios gives a €5bn bank the same vote as a €2.6tn one.

Read it
Supply chain · 10 September 2026

Late against what?

Three years of a retailer’s shipments and a late rate of 57%. The premium mode is late on every order, and the dearer of the other two delivers no faster than the cheaper one — only the promise differs. An on-time rate measures the promise as much as the delivery.

Read it
HR analytics · 10 September 2026

Counting people is harder than it looks

Five years of HR records and three rules that decide whether a headcount is right. The status field says 2,613; the dates say 1,480. Headcount is a stock and never sums, and a year that holds seven months has to be annualised over seven.

Read it
Data quality · 9 September 2026

What public datasets don’t tell you

Six datasets, six Power BI builds, and a pattern. None of them was dirty in the way people mean when they say dirty data — but every one contained something that changes the answer and announces itself nowhere. A key that is not a key, a status that contradicts its own dates, one column holding two different meanings, and an average that describes no one.

Read it
Data modelling · 9 September 2026

The same data, modelled twice

Everyone tells you a star schema is faster than one flat table. I built both over the same 1.7 million transactions and measured it, and the usual claim turns out to be the least interesting thing that happened. The flat model is six times the size, it is barely slower — and it answers one ordinary question wrong by a factor of seventeen, with no error and no warning.

Read it
Delivery · 9 September 2026

A Power BI report you can review

A PBIX is a zip file full of generated JSON. You cannot diff it, you cannot review it in a pull request, and you cannot test it — so the control on the report your board reads is usually a screenshot in a change ticket. Here is how these builds work instead, including the two bugs that got past every automated check.

Read it