Articles

Notes from the work

Practical writing on shipping AI and cloud systems that survive production. No think-pieces, no predictions — just the things that decide whether a project works.

Decision engineering: pointing abundant cognition at value

Agentic AI makes cognitive capacity abundant, so the constraint moves to knowing where to point it. Mapping the decision portfolio, the five rungs of the decision ladder, and a worked example from manufacturing master planning — down to a single configuration parameter with revenue on the other end.

How does AI generate value?

Six mechanisms turn cheaper cognition into money — substitution, augmentation, expansion, speed, consistency and new products. Which one applies decides whether your project pays for itself, and where the value leaks before it reaches the P&L.

How to calculate ROI on an LLM project without fooling yourself

The naive business case is always too high. A cost model that includes review labour, error remediation and maintenance, with the formulas, a worked example, and an interactive calculator you can run with your own numbers.

Building an LLM evaluation harness that catches regressions

Replacing “it seemed better in the playground” with a number you can defend: dataset design, graders from cheapest to most expensive, calibrating an LLM judge, handling noise, and gating CI on regressions rather than absolute scores.

Working on something like this?

These are the problems I get hired to solve. If one of them is on your plate, get in touch — a 30-minute call costs nothing and usually clarifies the scope.