Agents and automation
Loops that do work: read, decide, act through tools, and check. The stop rule is the part that makes it a product.
The loopAI and robotics
We are AI consultants and innovation consultants. We build the systems that answer for their output, and the cells that people can stand next to.
A model is easy to obtain and hard to be responsible for. Anyone can call an interface and get an answer back. What takes work is knowing whether the answer is right, what happens when it is not, and who is accountable when a customer acts on it.
That is the whole job. We take a question you have, find out whether a model is the right instrument for it at all, and then build the thing around the model that makes it usable: the evaluation set, the fallback, the logging, and the rule that stops it.
The answer is sometimes that you do not need a model at all. A lookup table, a threshold or a better form solves a surprising share of the work that arrives under that heading, and all three are cheaper to run and easier to defend in front of a customer.
Agents
An agent is a loop: something triggers it, it plans, it acts through tools, and it checks the result. The first three are the easy part and the demonstrations all stop there.
The check is where the engineering is. What counts as done, what counts as wrong, how many times it may retry, and what it does when it has run out of retries. A loop that cannot decide to stop will eventually spend your money or your reputation on a task it misunderstood.

Evaluation
A score on a public benchmark says how a model does on somebody else's problem. It is worth reading and it is not evidence about yours.
So the first thing we build is usually not a model. It is a set of your own cases with the right answers written down, including the awkward ones, and a way to rerun it. Once that exists every later decision has a number behind it, and you can rerun the whole thing yourself after we have gone.
The failures get enumerated before launch rather than discovered after. A failure you can name is a failure you can put a fallback behind.
Three subjects
Each of these has its own page, because the engineering in them is not the same.
Loops that do work: read, decide, act through tools, and check. The stop rule is the part that makes it a product.
The loopDetection with a confidence score, a threshold that somebody chose, and an honest account of what falls under it.
What it seesInference in the device, where the latency budget is fixed and there is no round trip to fall back on.
The budgetThe rules
European AI rules sort systems by what they are used for. The same model can sit in a minimal risk product and in a high risk one, and the obligations are completely different.
That is settled before launch, in writing, as part of the technical file. Doing it afterwards means discovering in a procurement questionnaire that the product needs a conformity assessment nobody planned for.
The declaration is the provider's own, which in almost every case is you and not us. We produce the evidence and the file it rests on, and we say plainly when a decision needs a lawyer rather than an engineer.
Robotics
A robot cell is designed from the safeguarded space inward. Reach, speed and what happens when somebody opens a gate come first, and the layout follows from them.
That is a whole subject and it has its own page.
Or write here
Name, an address we can answer to, and what you are trying to build. You get a reference back straight away.
Bring the question rather than the technology. We will tell you whether a model is the right instrument, including when it is not.