Your own AI model, running wherever you want it

The price depends on the scope of the task, the size of the model and the amount of data we work with – fine-tuning a smaller model for a specific task costs something different from training a larger model on domain-specific data. During the initial consultation we go through your brief and prepare an indicative tailored quote, so you know what to expect before we start the project.

We work with open language models (such as Llama or Mistral) as well as commercial APIs like GPT or Claude. Besides language models, we also train computer vision models and predictive models over business data. For more demanding training and fine-tuning we use the computing power of the LUMI supercomputer, one of the most powerful in Europe.

The model and all the data it is built on remain yours. We share nothing with third parties and never use it to train other projects – we work purely as a supplier on your infrastructure or in an environment we agree on in advance.

Yes, integration with existing systems is a normal part of our projects – whether it is an ERP, accounting software, an e-commerce platform or internal databases. The model can then work directly with your real data and fit into the processes you already have in the company.

First we clarify the goal together and go through the available data. Then comes data preparation and cleaning, the actual training or fine-tuning, testing on real scenarios and tuning based on the results. Finally we deploy the model and hand it over together with documentation on how to work with it.

A custom model makes sense for companies that work with specific or sensitive data, need to solve the same type of task repeatedly – such as quality control in production, document extraction or demand forecasting – or run into the limits of general tools like ChatGPT. If your task is more of a one-off or general nature, an off-the-shelf solution may be faster and cheaper – and we will tell you so honestly during the consultation.

We agree on regular consultations where we show interim results, discuss any changes to the brief and answer questions. If you prefer, we can also keep in touch continuously by e-mail or video call, whichever suits you.

Yes, there is no need to rebuild the model from scratch if your needs change or new data comes in. We can fine-tune it, extend it with new capabilities or retrain it on more recent data, so it grows with your company.

By agreement we can also run the model after deployment – including backups, monitoring and security measures. If you want to handle operation in-house, we hand over the model with documentation so you can manage it yourselves.

ChatGPT is an excellent general assistant, but many business tasks are of a different kind: you won’t solve defect detection on a production line, demand forecasting or document extraction with it. And even for language tasks, a custom or fine-tuned model has advantages – it keeps to your terminology and style, runs on infrastructure under your control, and at high query volumes works out cheaper than a paid API.

Data protection is essential for us – we work only with data you are authorised to use, and we set up the project so that sensitive data never leaves your infrastructure or the agreed secure environment. We comply with the GDPR, sign a non-disclosure agreement (NDA) with all parties, and never use your data to train other models or projects.

Yes, you don’t have to start from scratch. If you already work with a model and are not happy with the results, we look at exactly where it falls short and propose whether fine-tuning, adjusting prompts and data, or switching to a different model is the more effective route.

What helps most is an idea of the problem you want to solve and access to the relevant data – the documents, databases or systems the model is supposed to work with. We take care of the rest, including the technical setup.

Just write to us via the contact form on this page, call +420 555 333 158, or simply pick a slot for a 15-minute call in the calendar. We get back to you within one working day and on the intro call we discuss your brief, your data and the possible solutions. Neither the consultation nor the proposal costs anything or commits you to anything.

Additional training alone won’t; hallucinations are inherent to language models. Reliability is handled differently: by connecting the model to your documents and databases (RAG) so it answers from verified sources, and by thorough testing on real examples. Fine-tuning adds the right terminology and form of answers on top. We combine both according to what the task requires.

Yes. Language models are only part of our work. We also train computer vision models for quality control or reading documents, and predictive models for planning demand, stock and maintenance.

It depends on the size of the model and how many people will use it. For document data extraction and text classification a single powerful graphics card with 24 GB of memory is enough. For a company assistant over your documents a 32 GB card is a sensible start, and for several users at once 96 GB and up pays off. We never give the recommendation straight from a table though – we measure the model on your data first and propose the setup that genuinely covers the task.

The cloud is billed by the volume of text processed, so cost grows as AI settles into the company. Your own server is bought once and then costs electricity and administration. The turning point usually sits where the monthly API spend would cover the hardware payment, which for an agent running non-stop is a matter of months, while for occasional use it never arrives. The calculator on this page gives you the rough figure; the exact one comes from your real query volume.

On the most demanding tasks yes, and there is no point hiding it. The gap is far smaller than it used to be, and for everyday company work – data extraction, summarising, answering over your own data, classification – a well-deployed local model is fully usable today. What decides is not a leaderboard but whether it handles your specific task, which is why we measure it before deployment.

Quantization shrinks a model by storing its internal numbers at lower precision. A model with thirty billion parameters takes about sixty gigabytes at full precision and just under twenty at four-bit, so it fits on a single graphics card. The quality loss at four bits is usually small and at more aggressive levels noticeable. The exact figure for your task cannot be estimated, only measured on a test set, and that is part of our work.

Often yes, especially if it holds a graphics card with enough memory. We audit what you have, measure what will actually run on it and how fast, and only then say whether adding memory, swapping the card or buying a new machine is the answer. Sometimes topping up the existing server works out cheaper than a new one.

Yes. Once installed, a local model needs no outbound connection, which suits production floors, closed networks and field work. What remains is updates and backups, usually handled in a service window or on removable media according to your security rules.

A local model does not change by itself, which is its advantage – what you tested is what runs. New model versions appear continuously and moving to a newer one only makes sense when it measurably helps on your tasks. If you have operation and monitoring with us, we watch that for you and propose the move with numbers in hand.

Mainly it simplifies evidence. When processing happens on your server, there is no transfer of personal data to a third party and no question about where data is stored or for how long. The remaining obligations do not disappear, which is why deployment comes with access control, query logging and a description of how the system works, so there is something to build an audit on.

Yes, and it is one of the main reasons to go local. An agent running non-stop and processing every incoming e-mail or document consumes large volumes in the cloud and its bill is hard to predict. On your own model it runs for the price of electricity. Designing and building agents is covered by our AI agents and workflow service.

We build the solution so the model can be swapped without rebuilding everything around it. The interface your systems call stays the same, so you can deploy a larger model, move to stronger hardware, or route demanding queries to the cloud and keep the rest local. That combination is what most companies with a well-built setup end up using.