Run AI in your own environment when data can’t go to public AI services
Some data can’t go to a public AI service: client files, health records, contracts, source code. We build AI assistants on open-weight models (ones you can download and run yourself, such as Llama, Mistral or Qwen) inside your cloud account or on your servers. They answer from your documents and databases with a link to the source, and each person sees only what their role allows. If an EU-hosted business API already meets your rules, we’ll say so, because it’s cheaper and quicker to start.
The situations where teams usually call us.
Your security or legal team has said no to public AI tools for anything that contains client data.
Staff paste internal documents into public chatbots anyway, because nothing approved does the job.
You want to ask questions about contracts, policies or case files without uploading them anywhere.
A client contract or a regulator requires that data stays in your infrastructure or in one country.
You ran an AI pilot on a public API and now need something that passes your security review.
Everything here runs inside your environment. The model is only one part. Most of the work is finding the right documents and respecting permissions.
What we tell teams before they buy a GPU server.
Many data rules are met by a business API with EU data hosting, a data processing agreement and no training on your data. It’s cheaper and gives you the strongest models. Self-host when contracts, regulation or your own risk assessment say data must not leave your control, or when your volume is high and steady.
Open-weight models handle summaries, answers from documents, data extraction and drafts well. The best commercial models are still ahead on hard reasoning and long, complex tasks. Test both on your real questions, not on public benchmarks, before you decide.
Most poor answers come from the search step: scanned PDFs with no text layer, outdated versions, documents split in the wrong places. Cleaning up and structuring the sources usually improves answers more than switching to a bigger model.
A GPU server costs money whether it’s busy or not, and someone has to update the model, monitor it and patch the system. Renting GPUs in an EU cloud is often a better start than buying hardware. Buy once usage is steady and you know the size of model you need.
All three can be used safely for the right data. The difference is who processes it, and what you give up for control.
| Public AI API | EU-hosted business API | Self-hosted model | |
|---|---|---|---|
| Data stays in your infrastructure | No | No | Yes |
| Data stays in the EU | Partly | Yes | Yes |
| Strongest model quality | Yes | Yes | Partly |
| Time to start | Days | Days to weeks | Weeks |
| How cost behaves | Pay per use | Pay per use | Mostly fixed |
| Who maintains the model | Provider | Provider | You or us |
| Works without internet access | No | No | Yes |
A typical timeline. We confirm yours after the project workshop.
Week 1
Data rules and one use caseWe go through the data involved and what your rules and contracts require, then pick one use case with a clear owner, such as questions about contracts or internal policies. You get a fixed price.
Weeks 2–3
Test models on your questionsWe run open models on a set of your real questions next to an EU-hosted API, so you see the quality difference before choosing.
Weeks 4–6
Build and connectDeployment in your cloud or on your servers, document sources, permissions and logging. A working demo every week.
Then
Pilot and widenOne team uses it daily while we measure answer quality, then we add sources and teams. You own the code and the setup.
Systems we’ve built that are close to this one.
An AI assistant that answers questions about building permits and contractors using current database records.
Read the case study
Datair uses AI to help people analyze data and create reports without technical knowledge.
Read the case studyNot to start. A pilot can run on GPUs rented in an EU cloud, in your own account. Buying hardware makes sense once usage is steady and you know what size of model you need.
For summaries, answers from documents, extraction and drafts, current open-weight models are good enough for most internal use. On hard reasoning and long, complex tasks the best commercial models are still ahead. We test both on your real questions so you can see the difference before you decide.
A private assistant with several sources and permissions usually fits our connected-system range of €10–30k. A pilot on a single source can fit the €5–10k range. Hardware or rented GPUs are a separate running cost, which we estimate before you commit, and the build price is fixed after the project workshop.
It removes one big question: which outside company processes your data. GDPR still applies to what you store and for how long, and the EU AI Act adds duties depending on what the system is used for. We help document the setup, but it isn’t legal advice, so your data protection officer or lawyer should confirm it.
Yes. It checks who is asking and searches only the documents and records that person can open in the source system. Someone in sales won’t get answers from HR files.
Usually not. The model reads the relevant passages when a question is asked (RAG), which is cheaper, keeps answers current and makes deleting data simple. Fine-tuning pays off only in narrow cases, such as a fixed output format or a specialised classification task at high volume.
You own the code, configuration and accounts. Bug fixes are free for 3 months. After that we can handle updates and monitoring on monthly support, or hand everything over to your IT team.
A first version usually takes 6 weeks to 3 months, depending on how many sources are involved and whether hardware is already in place. Starting on rented GPUs avoids waiting for hardware delivery.
Tell us what data is involved and what your rules require. We’ll tell you whether you need self-hosting at all, and what a first use case would cost.
JurajCEO, QikBuild
You leave the call with
Or write to hello@qikbuild.com
