About
I'm Samuel McDonnell, an ai engineer at RWA in Ireland. I work on the part of AI systems that decides whether they can be trusted in production: the evaluation harnesses, the tool boundaries, the budgets, and the measurements that tell you whether a change helped.
What I build
Most of my work is making model behaviour legible. That means typed boundaries between a model and the systems it can reach, so a bad call fails as a validation error instead of a mystery downstream. It means hard caps on tool calls and wall-clock time, so an agent's cost is a known quantity rather than something you discover at the end of the month. And it means scoring — rubrics written down before the run, so a change in output can be attributed to the world changing or to me changing the bar, and not to the model having a different day.
The through-line is that a system you cannot measure is a system you cannot improve, and one you cannot honestly recommend.
Why this site is shaped like this
Every project page opens with either a number and the method behind it, or a written statement of why no number exists. That constraint is enforced by the content schema — the site will not build a project that has neither. It is the same discipline I'd want from anything I shipped, applied to the thing that describes what I shipped.
Most of the projects here currently sit on the second branch. That is the honest state of them, and stating it plainly costs less than a number I could not let anyone check.