AI & machine learning
Not a general-purpose AI with your logo on it, but a model that has learned from your own photos, figures or planning rules. Three kinds, all three trained by me. A fixed price agreed up front, and it always starts with whether this can work in your case at all.
Two half-days walking through your work: what data do you have, what can realistically be done with it, and what would it be worth? You get a report, not a model.
€495
Lead time: 1 to 2 weeks
A model that learns to sort, inspect, count or recognise from your own photos. Set up as a trial project, ending with a measurement on images the model has never seen.
from €2.950
Lead time: 4 to 16 weeks
A model that works out the best arrangement within your rules: who works when, which order goes on which run, how much of what to buy. Not a forecast but a decision, and one question at a time.
from €2.450
Lead time: 4 to 14 weeks
Customer churn, how busy each week will be, no-shows, lead times or price sensitivity. Works on the figures already sitting in your bookkeeping or point-of-sale system.
from €1.950
Lead time: 3 to 13 weeks
A model is not a website. It quietly gets worse as the world changes and the model does not, and it does so without an error message. This keeps it running and measured.
€99 to €249 per month
Lead time: ongoing
What a model costs depends on how big the question is; the amounts above are the smallest setup. In the configurator you pick the size and the amount follows along.
All amounts exclude VAT.
Not sure which way to go yet? Start with the scan: it costs the least and prevents the most expensive mistake. Or start with what I do in web design.
I built a classification model that had to tell from photographs whether something passed or failed, and what that project mostly taught me is where this kind of work comes apart in practice. Those lessons are worth more than the list of techniques underneath them, so here they are.
The model that scored best was the worst model. Training automatically picked the version with the lowest error, and that turned out to be a model that simply approved everything. Beautiful on paper, catches nothing in practice. Since then I never look at a single number, but always at what goes wrong and what that costs.
The measurement was an illusion at first. Photos from the same batch sat in both the practice set and the test, and then you are measuring how well the model remembers rather than how well it recognises. It is the most common mistake in this field and it is invisible from the outside: the figures look too good, not wrong.
And there is nearly always a dial that costs no retraining. Where you draw the line between pass and fail decides whether you catch more faults or throw away more good product. That is a business decision, not a technical one, and it is exactly the conversation I want to have with you before anything gets built.
The first question is always whether it can be done with a handful of examples. For image recognition and forecasting: it cannot. Count on hundreds of examples per category, and that applies just as much to the category you are actually trying to catch. Twenty photos of a defect is too few to teach a model what a defect is, however good the rest of your data may be.
Beyond that you need someone who can say what is right and what is wrong. A model learns from that judgement, so if two colleagues disagree about the same photo, that is a problem to settle before training rather than during it. That costs time on your side, and it goes into the quote up front instead of halfway through.
With an optimisation model it is the other way round. That one learns from rules rather than examples, so the question is not how much data you have but whether someone can write down what a good plan has to satisfy. That sounds easier and usually is not: half of those rules live in the head of whoever does it now, and only come out once the model proposes something that plainly cannot be done.
And this is research, not a build. With a website I know in advance that it will exist and the only question is how. Here, “this cannot be made good enough” is a valid outcome that you agree to up front. That is why the scan exists, why the image work is a trial project, and why you will hear a range from me and never a percentage before I have seen your data.
Your data goes to a server I run myself, in the Netherlands. That is not a marketing line but the difference with nearly every provider in this segment: they send your documents and photos to a service in the United States and call it bespoke.
No external AI service is involved. The three models above are trained by me on your material and then run on that same server. If something from outside were needed in your case after all, you hear it before it happens, it appears by name and place of business in the privacy statement and the processing register, and we sign a processing agreement.
Whatever you hand over for training stays yours. I do not use it for other clients and not to improve anything general, and afterwards you get it back or it is deleted. That sits in the contract, not only on this page.
No mounting cameras on a production line. As soon as lighting, hardware and a link to the line controls come into it, you are talking about tens of thousands of euros and about maintenance on site. That is work for a firm with engineers, and I would rather tell you straight away that I am not one.
No decisions about people. Medical assessments, creditworthiness, pre-selecting job applicants: those uses fall partly into the high-risk category of the AI Act, with obligations a one-person business cannot carry and consequences you do not want to explain when it goes wrong.
No chatbots and no language models. That is the busiest corner of the market and the easiest to walk into: a key with an American provider, a few documents dropped in, and you have a product. Except then you are not training a model, you are building a shell around someone else's, and everything above about where your data stays no longer holds. Hence three models I train myself, and nothing besides.
And no promising accuracy before I have seen your data. Anyone who gives you a percentage up front is selling you a guess with the authority of a number.