About
I am Richard Atkins. I build AI systems for a living and I teach the part that survives contact with real work.
Twenty-plus years as a tech lead and solution architect, through my own consultancy, for insurers, banks, automotive, SaaS and media. For the last few years the work has been AI: agents, retrieval, evaluation suites, cost models, and the organisational side of getting any of it adopted.
I run production AI at Cedar & Bloom, which is where the ladder comes from. My agents are promoted from shadow to autonomous on evidence, and demoted when the numbers slip. I write about what I measure under the name The Grounded Architect, and I publish the part where I was wrong.
The teaching started by accident. Friends and colleagues asked how I used these tools, I showed them on their own work, taking them from zero to daily use.
I live between Cambridge and Zurich, which is why onsite means the UK and everything else is remote in UK hours. My other site, with the case studies and the long CV, is at richard-atkins.com.
Evidence
What I can point at.
Production AI, not slideware
I build and run AI systems at Cedar & Bloom: an AI news platform and a careers service, with agents, retrieval, evaluation suites and a measured cost model. When I teach a method it is because I have run it and watched it fail first.
Open-source reference repos
An MCP server that ships with its own eval suite. An agent harness where every run is judged. Evals as a CI gate, a resumable cost-aware pipeline, and two habit analysers that refuse to report on thin evidence. Six repos, all MIT-licensed, all yours to keep after a session.
A dozen published field notes
Engineering pieces on what survives production, and five essays on what AI does to the people around it. Every number in them is measured or labelled as modelled. Honesty over hype is the whole editorial policy.
Twenty-plus years of delivery
Tech lead and solution architect for insurers, banks, automotive, SaaS and media, through my own consultancy. I have been the person in the room who had to make the plan work, not the one who presented it and left.
Open source
Reference repos you keep after a session.
MIT-licensed, each built to prove what it claims. Most have a companion field note covering what was measured and what it cannot do.
What I will not do
Five things, so you know before you book.
- Sell you a tool. I do not resell anything and I have no partner deals.
- Promise a multiplier. Anyone quoting one has not measured it on your work.
- Teach a method I have not run in production and watched fail at least once.
- Put your data into a model without your say-so.
- Tell you the climb is worth it when it is not.
Tell me which rung you are on.
A short message is enough. Say who you are, what you are trying to do, and whether you want remote or onsite. I reply within two working days.