Can we train a model on our company data without it leaving?

Yes. We start from an open model and fine-tune it on your infrastructure, or in a private European cloud under your control. The data is never sent to a model provider, neither for fine-tuning nor for answering.

Is a local LLM as good as ChatGPT?

For chatting about anything, no. For a specific task on your documents —search, summarise, classify, extract data— a well-chosen open model is usually enough. We check it with your own questions before deciding, not with a generic demo.

What hardware do we need?

It depends on the model size and the number of queries: from a workstation with a graphics card to a dedicated server. It's estimated in the diagnosis, with the cost and payback period alongside, before anything is bought.

Is a local LLM GDPR-compliant?

It makes it much easier, because there's no transfer to third parties or outside the EU, but it doesn't exempt you from anything: you still need a legal basis for processing and proper security. We design for that from the start.

KALACALA · Creative Technology

KALACALALet's talk

Service — Private AI and custom modelsWhen the data can't leave

Private AI that stays in-house

Open models, fine-tuned on your company's data and deployed on your infrastructure —or in the browser itself— for when information can't leave. No documents sent to third parties, and no per-query bill.

Terms

Where it runs: On your servers, in a private European cloud or on the user's own device.
Data: Stays put: not for fine-tuning, not for answering.
Price: Per project, fixed in writing. The hardware needed is estimated before we start.

Tell us about the data

Some information can't travel: client files, medical records, drawings, formulas, contracts. For those companies the question isn't which model is best but which one can be used without the data leaving the building. In 2026 the answer is a good one: open models handle specific tasks —search, summarise, classify, extract— well, running on your own server.

What you shouldn't do is promise more than they give. A local model doesn't write like the best commercial model, and it doesn't need to: answering from your documents doesn't require it. We measure with your questions before deciding, and if a task needs an external model, we say so.

Who it's for

Law firms, accountants and notaries working under professional secrecy.
Healthcare, labs and insurance, handling special-category data.
Manufacturers with proprietary know-how —drawings, formulas, processes— that can't end up with a third party.
Companies whose clients contractually forbid sending their data to external services.

What we build

On-premise deployment

The model on your infrastructure. On your own server or in a private European cloud, with your systems' permissions and no per-query fee to any provider.

Custom models

Fine-tuned on your data. We start from an open model and fine-tune it on your documents and vocabulary for a specific task. The resulting model is yours.

Private search

Over internal documents, with answers citing the document they come from, and no content passing through any external service.

In the browser

AI running on the user's own computer. For tools where not even the server should see the file: processing happens on the device, with WebGPU or WebAssembly.

Sizing

How much hardware, and whether it pays off. Which machine the real volume needs, what it costs and how long it takes to pay back against pay-per-use. Sometimes the answer is that it doesn't, and we say so.

Evaluation

The same control questions as any AI system: we measure whether the local model gets right what it has to before it goes into use.

What you get

The system running on your infrastructure, with no external dependency to answer.
The fine-tuned model, if there is one, with its training data documented. It's yours.
Accuracy measured against your control questions.
The hardware and running-cost estimate, before anything is bought.

Where it shows

SecureConverter

SecureConverter processes video, audio, images and documents with AI that runs entirely in the browser: 0 file bytes sent to the server, across three execution engines depending on the device. Read the case →

What people ask us

Can we train a model on our company data without it leaving?: Yes. We start from an open model and fine-tune it on your infrastructure, or in a private European cloud under your control. The data is never sent to a model provider, neither for fine-tuning nor for answering.
Is a local LLM as good as ChatGPT?: For chatting about anything, no. For a specific task on your documents —search, summarise, classify, extract data— a well-chosen open model is usually enough. We check it with your own questions before deciding, not with a generic demo.
What hardware do we need?: It depends on the model size and the number of queries: from a workstation with a graphics card to a dedicated server. It's estimated in the diagnosis, with the cost and payback period alongside, before anything is bought.
Is a local LLM GDPR-compliant?: It makes it much easier, because there's no transfer to third parties or outside the EU, but it doesn't exempt you from anything: you still need a legal basis for processing and proper security. We design for that from the start.

How it's bought

Every service follows the same path: a short diagnosis, the build with weekly deliveries and, if you want it, ongoing maintenance.

Related services

Let's talk

Tell us what you're building.