Claims about AI stream towards a test on real tasks. Most are stopped. The ones a test backs come through with their measured result: 17 of 30 real coding tasks solved, 38.9 tokens a second, and running on local hardware with no cloud.
ResearcherAuthorTester of open AI models
What does AI change for the people who decide?
Nearly two decades. Four countries. Five fields.
Every month brings new claims about AI, and few of them have been tested on real work. I have spent nearly two decades between the people who decide and the systems they decide with, and I now test AI the way an engineer tests a machine, so that you can see what holds up before you trust it.
Collaborated with 35 DAX-listed · 12 FTSE 100-listedRegistered in the European Commission Expert Database
Learning with my hands first.
As a boy I took things apart to see how they worked. That is how I decided to become an engineer. My first qualification was a diploma in marine engineering. In the evenings I added one certification each semester (arc welding, foundry work, refrigeration and air-conditioning repair, as well as three CAD systems), thus learning that knowledge lives in hands as much as in manuals. Then I started at a shop counter. Selling taught me the customer before the system. Later, a BSc (Hons) in European Engineering Business Management at Coventry University put the engineering and the commerce into one frame.
From the counter, the work moved across four countries.
Where the decision is made.
Over nearly two decades I worked in the UK, France, Hong Kong and Germany, in five fields: retail and e-commerce, marketing and media, supply chain and logistics, industrial and energy, and now AI and new technology. The roles grew from sales into strategy and consulting, including a role as CTO, carrying the technical and the commercial side together. The industries changed, but the pattern stayed the same. A decision can only be as good as the knowledge that reaches the room, and that knowledge often sits with people who are not in it.
Until one day I decided to study that pattern properly.
Who shares what they know?
I saw it in every field, so I made it my doctorate. The MBA at the University of Bolton (with The Bolton Award) had already taught me to test theory against real work. The doctorate, completed at Manchester Metropolitan University, studied knowledge-sharing behaviour: how people perceive sharing, what they intend, what they actually do, and what makes them share or hold back. In my reading, those patterns are what AI now runs on. Models are trained on what people chose to share, use it to answer, and pass it on to whoever asks. That is where AI's influence on organisations comes from, and also where it can quietly thin out the human expertise it draws on. Neither can be judged from a vendor's slide.
So I began testing the tools myself, on real work.
Every model earns its place.
You only understand a machine once you have taken it apart, and AI is no different. I run open models on my own machines, with cloud models beside them, and test them on real work such as fixing software issues and pulling fields out of documents. Not every model stays. Nemotron 3.5 Lightning, a 30-billion-parameter model, solved 15 of 30 coding tasks where a faster model solved 18. On one task it produced ten times as much text to reach an answer. I removed it. I also fine-tune my own Kev checkpoints, following Jared Palmer's open-source recipe on an open base model. A plain test, not a model, decides whether the work passes. That is the rule of the whole lab: a local model can fail work, but it can never approve it.
What held up across those years of work, I have written down.
The method, written down.
Part of it is already in print: two chapters I wrote with Farag Edghiem for an IGI Global book on digital entrepreneurship. Four finished manuscripts hold the rest of the method, on decisions, communication, productivity and habits. Each one is a set of compact frameworks that a reader can use at work the next day, thus turning what was tested in the field into something other people can repeat.