Notes
Writing that came out of the work.
Most of what I write lives next to the code it describes: investigations, model cards and reports that explain what was built, what was thrown away and why. The best of it is collected here.
(01) — The index
01Machine learningAugust 2026
Why the first model was thrown away↗
Three model families scored a perfect ROC-AUC of 1.0000. The investigation into a missingness pattern that matched the target on all 148,670 rows, and what the honest number turned out to be.
02FairnessAugust 2026
What it costs to exclude gender and age from a credit model↗
Four and a half basis points of ROC-AUC. A fair-lending analysis that measures outcomes across groups the model never sees, using the adverse impact ratio as a review trigger rather than a legal test.
03DocumentationAugust 2026
Model card: mortgage probability of default↗
What is modelled, what is assumed, and what the model must not be used for. The most important distinction on the card is between the PD that was estimated and the LGD that was not.
04RetrievalAugust 2026
Embeddings from different models are not comparable↗
Query an index built with one embedding model using vectors from another and nothing fails. You just get confidently wrong passages. One namespace per model, and two details that must never drift apart.
05EconomicsMay 2026
Nigeria's crude-oil exports, 2010–2024↗
A trade-intelligence report: price elasticity of 0.74, the COVID collapse as the only true statistical shock, a post-2021 volatility regime, and why AfCFTA has not touched intra-African oil trade.