I help companies get from a pilot that impressed the board to a system you can trust with real customers and real data. I train teams and build AI systems, and I check how safe they are. For larger projects the Prosit team works with me.
PLN 7,000 net per consulting day · projects quoted after the audit
One process handled by AI in production, with numbers that show it works.
We start with the AI decision audit: which decisions a model can take over, tested on your data.
A pilot on one process, then production with monitoring, evaluation and escalation to a person.
Local and sovereign models, such as Bielik, when your data cannot leave the company.
Reference: Chatmerce, our RCS, SMS and web agent platform, where escalation is a real tool, test suites run on every change, and an LLM judge feeds retraining.
PLN 7,000 net per consulting day · projects quoted after the audit
You know where your LLM app or agent can be attacked, what to fix first, and which regulatory deadline applies to you.
Security review of LLM apps, agents and MCP servers against the OWASP Top 10 for LLM Applications (2025) and the OWASP Top 10 for Agentic Applications (2026): prompt injection, tool permissions, data leaks.
Secure development with AI coding assistants: secrets, hallucinated package names (slopsquatting), SBOM and supply chain, NIST SSDF.
Vulnerability monitoring that sets priorities (CVE, CISA KEV, EPSS, VEX). I designed and lead development of Siemens Energy's in-house industrial cybersecurity portal: OT and IT asset inventory, risk, vulnerabilities and patch guidance.
Regulatory readiness. Since 11 September 2026 the Cyber Resilience Act requires manufacturers to report actively exploited vulnerabilities within 24 hours. The amended Polish KSC act (NIS2) has applied since 3 April 2026, with security measures due by 3 April 2027. Also DORA, the AI Act and IEC 62443 for industrial systems.
Prices are net, VAT is added. I issue a VAT invoice, in EUR or NOK on request. I can sign an NDA before our first call.
Proof
Three days with PKO Bank Polski developers
20 / 20participants deployed a working app on day three
3.7 → 7.6average knowledge test score, out of 10, before and after
33% → 2%share of “I don’t know” answers in the same test
“AI in software development”, a three-day workshop run through Altkom Akademia, September 2026.
Research on when a model should say “I don’t know”
Two papers on Polish Bielik models: a probe on the model's activations reads that a name is unfamiliar before the first word is generated, and the signal is graded, so it can drive a decision to answer, search or hand over to a person.
Industrial cybersecurity portal for Siemens Energy
I designed and lead development of the company's in-house portal for managing industrial cybersecurity: an inventory of hardware and software across OT and IT, threat and risk management, vulnerability tracking and patch guidance. Built on Go, GraphQL, React, PostgreSQL, Elasticsearch and ClickHouse, running on Kubernetes in AWS.
Agents answering customers in production
At Prosit we built Chatmerce, an agent platform for RCS, SMS and the web. It hands conversations to a person when it should, and every change goes through test suites and an LLM judge before it reaches customers.
Talks
AI Miners #7 in Katowice (September 2026), an AI Ninjas workshop on proving that AI wrote a text (August 2026), and Eskadra Bielika in Gliwice, where participants built a RAG system on a Polish model (June 2026).
Some background: I am CTO and co-founder of Prosit, a Polish-Norwegian consultancy of 16 people that works for clients such as Siemens Energy, Grønt Punkt Norge and Handelens Miljøfond. I trained as a physicist at the Jagiellonian University.
How I work
A free 30-minute callYou tell me what hurts. No sales deck. If I am not the right person, I will say so and point you to someone better.
DiagnosisI look at your code, your process or your data, depending on the problem, and talk to the people who do the work.
A plan you keepA written plan with priorities and cost. It stays with you even if we go no further.
Training or implementationI run the workshop or build the solution with your team, with Prosit engineers when the scope needs more hands.
Handover and supportDocumentation, tests and the scripts to rerun them. I am done when your team manages without me.
Questions
How much does AI training for a company cost?
A training day costs PLN 7,000 net. A typical workshop takes one to three days. Group size, location and travel we settle on the first call, and you get a fixed price before anything starts.
Can the training use our own code? Will you sign an NDA?
Yes to both. Exercises on your own repository work best, because people learn on problems they will meet on Monday. I sign the NDA before I see any code, and we can run everything on your machines and your licences.
Where does our data go?
Wherever you decide. Local models such as Bielik run on your own hardware, so data never leaves the company. When a cloud model makes sense, we use providers and regions in the EU, and only for data you are allowed to share.
We already have Copilot. Why would we need training?
A licence does not change how people work. At PKO Bank Polski the average knowledge test score went from 3.7 to 7.6 out of 10 in three days, in a team that already had the tools. The difference comes from instruction files, custom agents, review habits and knowing when not to trust the assistant.
Do you run security tests of applications with LLMs?
Yes. I review the architecture and the permissions of tools and agents, then try to break the system with prompt injection, tool abuse and data exfiltration attempts. You get a report with priorities and, if you want, help with the fixes. For a broader scope, such as infrastructure tests, the Prosit cybersecurity team joins.
Do you work in Norway and in English?
Yes. I split my time between Rybnik and Oslo, Prosit is a Polish-Norwegian company, and I run training and projects in English as often as in Polish. Invoices can be in EUR or NOK.
Can the training be co-funded?
Ask me on the call. I will tell you what I can prepare for your application.
Tell me where it hurts
Thirty minutes is usually enough to tell whether I can help and what the first step would be.
