Twelve Consulting — on assignment at Rothschild & Co
This site is a notebook, not a shop window. I publish what I understand about work in the age of language models: what holds up in production, what breaks, and what it changes for teams. I write to think clearly — if it helps you too, all the better.
Almost everyone can point to an AI tool that saves them time. Far fewer can say whether they've become more capable because of it — or just more dependent on it. A case for AI that builds skill instead of quietly eroding it.
The projects I share openly, with their code and their reasoning.
Learn to evaluate products that embed AI — from your first failing run to a complete evaluation package you could hand to a decision owner.
Testing an AI product has little to do with testing classic software: several answers can be acceptable, and exact comparison no longer holds. This course does not benchmark models; it teaches you to define what "trustworthy" means for a given product, then to prove it. Seven modules, eight solved exercises and a capstone, built for non-engineers who have to sign off an AI system before it ships.
Three formats, one rule: never write anything I do not actually believe.
Long-form thinking: AI at work, product leadership, technical learning. Opinionated, and dated on purpose.
Short form: a tool, a read, a technical find, and why it is worth your attention.
Running, work habits, side quests. What fits no professional box but shapes everything else.
What I am working on today, updated when it changes.
Use-case scoping, roadmap, answer-quality evaluation (LLM-as-a-judge), guardrails, and GDPR / AI Act compliance. The interesting part is never the model — it is everything you must build around it before anyone can trust it.
The goal is explicit: not just to lead agentic projects, but to build them myself, locally. A product owner who can read and write code makes sharper calls — and gets told fewer stories.
A public repository helping non-engineers test AI products properly. Because "it seems to work" is not a test protocol.
View the repositorySixteen kilometres and a hill everyone dreads. I am training, taking notes on what works, and it will probably end up in the Personal category.
From frontline sales to AI product leadership, by way of a startup built from nothing.
A well-argued disagreement beats a polite agreement. If a post makes you react, or you want to stress-test an idea, write to me.