Custom AI models
Not every problem is solved by a better prompt. Sometimes you need a model that understands your data specifically.
When companies call us
- A general model does not understand your terminology, formats or domain context.
- Data must not leave the company — legislation, a customer contract or an internal rule.
- You need to process images: quality control, detection, measurement, satellite or drone imagery.
- At high volumes, token costs grow faster than the benefit.
This is where our roots run deepest. Patrik holds a PhD in artificial intelligence, has worked on computer vision from research to real-world deployment, and has led teams on projects where a model has to hold up on data it has never seen — including for international clients.
In practice this means we can not only call a model through an API, but also train, evaluate and deploy one — including training on HPC infrastructure and running it on Kubernetes.
At the same time we will be the first to tell you that you do not need a custom model. In most cases starting with a good prompt and RAG is cheaper and faster. A custom model makes sense for a narrow domain, strict data requirements or high volumes.
How we work
1. Verify a custom model is needed at all
We compare a baseline using an existing model against the expected gain from training. If the gap is small, we save you tens of thousands of euros.
2. Data and annotation
Dataset quality matters more than the choice of architecture. We help design the annotation process and check consistency.
3. Training and evaluation
Fine-tuning or training from scratch depending on the task. The result is always compared against the baseline on a held-out test set.
4. Deploy where it needs to run
In the cloud, in your data centre or at the edge. Including latency and inference-cost optimisation.
What you get
- A trained model with a documented, reproducible training pipeline.
- A comparison against the baseline on a held-out test set — including the cases where the model fails.
- Production deployment (cloud, on-prem or edge) with monitoring.
- The model weights, the data and the code are yours. No vendor lock-in from our side.
- An inference cost estimate at your expected volume.
Indicative scope
- Feasibility study and baseline
- €1,500 – €3,000
- Fine-tuning an existing model
- €4,000 – €12,000
- Custom computer vision model
- from €12,000
- On-prem deployment and operations
- individual
Prices are indicative and exclude VAT. Training cost depends mainly on the state and volume of annotated data.
Frequently asked questions
When does a custom model genuinely pay off?
When you have a narrow domain with its own terminology, a high volume of similar requests, strict requirements on where data may reside, or a need for low latency. If none of these apply, start without one.
How much data do we need?
For fine-tuning a language model, a few hundred to a few thousand good examples is often enough. Computer vision typically needs an order of magnitude more and depends on scene variability. We give an exact figure after looking at your data.
Can we run the model without an internet connection?
Yes. That is one of the main reasons companies reach for a custom model. Deployment on your own hardware or at the edge is routine for us.
What happens when a better model comes out next year?
That is why we keep training reproducible and maintain a test set from the start. Moving to a newer base is then a matter of retraining and comparing numbers, not starting over.
We will tell you straight whether your problem needs a custom model — even when the answer is “no, save your money”.