Questions and answers

Three things, layered. AI infrastructure is the platform underneath every AI feature — model routing, retrieval, evaluation, observability, cost control. AI agents are systems that take actions inside your tools with proper permissions and human oversight. AI chatbots on websites are the customer-facing version, in English, Czech and Slovak, instrumented so you can prove whether they help.

Look at where you are stuck. If the pilot works but nobody can deploy it or measure it, you need infrastructure. If the model cannot do the job because it needs to touch real systems, you need agents. If visitors are emailing the same eight questions, you need a website assistant. If you are not sure, that is exactly what the audit is for.

A focused website assistant starts around €4,900. An agent prototype in shadow mode is €6,500–15,000. Platform and infrastructure work runs at €850–1,100 per day. Ongoing operation is €1,200–3,500 per month. Full ranges are on the pricing page.

Audit 1–2 weeks. Foundation 2–4 weeks. First production feature 3–8 weeks depending on scope. Speed is mostly a function of how clean your data is and how many approvals a release needs, not of how fast I type.

The one that performs best on your evaluation set at an acceptable cost — which is usually not knowable before you build the evaluation set. My default is Anthropic for complex reasoning, OpenAI where you need the widest ecosystem, and smaller models for classification and routing.

Usually not. Hosted APIs plus a well-built gateway are cheaper and faster to start for almost every enterprise use case. Self-hosting becomes the right answer when data cannot leave your network, when you have steady high volume, or when a specific open model measurably beats hosted models on your workload.

Retrieval for anything that changes; fine-tuning for behaviour, style and format. Fine-tuning will not teach a model your updated refund policy. The longer argument is in RAG vs fine-tuning.

With Anthropic, OpenAI and Google enterprise terms, your data is not used for training by default. That is a contractual point, and I will make sure the specific tier you sign matches the specific promises you need. If you need certainty beyond a contract, the answer is self-hosted inference.

Yes, and I prefer it. Python, Java, .NET, anything — the AI layer will match your team rather than the other way round. I will not rewrite working infrastructure to suit my preferences.

It has to, to be useful — but the design controls that. Data stays in your infrastructure, travels encrypted, is masked at the boundary where it needs to be, and is not written to third-party logs. Access to the knowledge base is filtered per user, at query time, using your own authorisation.

It can be, with deliberate work: no third-party tracking, consent before storing personal data, short retention for transcripts, IP addresses not logged, processors listed in your privacy policy, and a clear note that the assistant is an automated system. I will also tell you which parts you cannot honestly make compliant.

By not trusting content or prompts by default. Untrusted text is isolated from instructions, each tool carries its own minimal permissions, outputs are validated, and there is an adversarial test set that runs in CI. Expecting a prompt to be injection-proof is the mistake I try hardest to prevent.

Yes. Mutual NDA before any system detail, and I am used to security questionnaires, DPAs and data-processing agreements. Where requirements are strict, I can deliver in a fully on-premise environment.

Almost always. Remote within the EU is the default; on-site in Bratislava or Vienna by arrangement. I have worked with teams in the US and Asia by structuring around written updates and async decisions.

That is my preferred model. I work inside your repositories and your review process, and I leave behind documentation and tests rather than a black box. Several of my engagements have ended because the team no longer needed me.

Then you find out in the audit, for a fixed fee, and you keep the plan. In delivery phases, acceptance criteria are written down before work starts, so "it does not work" is a defined conversation rather than a disagreement.

English, Czech and Slovak — for communication, documentation and products. Every multilingual system I build is designed for all three from the start, including separate evaluation sets per language.