Silence is not
the same as health.

Systems that run without anyone watching tend to fail quietly. I build them, work out how to tell when they have stopped, and write up what I find.

Scheduled agent · last 24 runsevery 15 min
did work ran, nothing to do did not run

Built

Game Oh! My Dog! A vet game where the finding that cracks the case is free and the expensive test only confirms it, built on a suite that once passed every check on a patient who could not exist. Tool stillrunning A dead man's switch for unattended jobs. Each one declares how often it will check in, and the alarm is that the check-in never came. Product Bitwright Pixel art described as data rather than generated as an image, so fifty sprites drawn one at a time still look like one artist made them. Tool Hum Music described as constraints instead of notes, rendered to separate layers a game fades in as things get worse. Open source · MIT Model Arena One prompt, sent to every provider at once, with the answers side by side and what each one really cost. Product Invoice Badger Invoicing for UK freelancers that you drive by talking to your AI, with a web UI for everything the conversation is bad at.
All projects →

Recent findings

Writing · 17 August 2026 A job that runs on time and does nothing Building a dead man's switch for unattended jobs, and the two bugs that nearly shipped in it. Both were the monitor inventing certainty it did not have, in opposite directions. Writing · 17 August 2026 What the judge does instead of judging I scored eighteen model answers by hand and compared myself to an LLM judge on the same rubric. It agreed with me 94% of the time on the checks and 39% on the one criterion that needed judgement, where it reached for answer length instead. Writing · 14 August 2026 When the model keeps asking for the same thing A tenth of my draws were failing. The model was not choosing badly, it was asking for something the format could not express, and rewriting the prompt made it worse.
All writing →

What I work on

Infrastructure and platform work

Cloud, networks, databases, deployment, the on-call rota. Twenty years of it, including the unglamorous parts: cost and capacity, and the migrations nobody wants to own.

AI agents in production

Scheduled jobs and AI agents that act without a person watching. Blast radius, recovery, and telling the difference between quiet and stopped.

LLM evals and cost control

How you find out a model output got worse after somebody else changed something, and what it costs per completed task rather than per token.

AI governance, written by the engineer

AI registers, DPIAs for live systems, EU AI Act Article 50 obligations. Most people in this market are lawyers who cannot read a config file. I work on both sides.

Who

Twenty yearsthe last few, mostly AI

Twenty years in infrastructure. Cloud, networks, databases, deployment, and being the person who gets called when it stops. Most of that has been keeping things running that other people depend on.

For the last few years a lot of it has been AI systems, mainly because those are the ones now running unattended. The problem underneath has not changed: knowing whether the thing is actually working, and finding out before somebody else does.

I write up what I find, usually because something looked fine and was not.

Get in touch

If something you depend on runs without anyone watching it, and you are not certain what would tell you it had stopped, that is the conversation I enjoy most.

hello@dmooney.engineer

Infrastructure, cloud and platform work as well as AI: what I can help with →

Reading