AI that runs on your own data — and complies with the law
I design and deploy AI systems for companies: search across internal documents (RAG), model fine-tuning and process automation. Models run locally — on your server or on my hardware in Poland. AI Act and GDPR compliance is designed together with the system, not bolted on after deployment.
Three things that pay off fastest in a company
I don't sell "digital transformation". I deploy concrete systems that save your team time — and I say plainly what today's models cannot do.
RAG — questions answered from your documents
A system that answers questions based on your contracts, procedures, technical documentation or ticket history — citing the source of every answer. Staff ask in plain language instead of digging through folders.
Fine-tuning — a model for your industry
Adapting a language model to the vocabulary and formats your company uses: product naming, document templates, industry jargon. The result: less manual correction and a predictable output format.
Process automation
Connecting AI to what your company already uses: email, spreadsheets, CRM, supplier APIs. Sorting mail, extracting data from invoices, drafting replies for human approval — with control retained at every step.
AI Act compliance — before it becomes a problem
Regulation 2024/1689 (the AI Act) applies across the EU, with obligations phasing in through 2028. They cover not only model developers but also companies deploying AI internally. Fines reach EUR 35 million or 7% of worldwide turnover. I help establish what actually applies to you and prepare the documentation before a client, auditor or insurer asks for it.
Risk classification
Determining which AI Act category your system falls into — prohibited, high-risk, limited or minimal — and whether you act as a provider or a deployer. The entire scope of obligations follows from this, and misclassification is costly in both directions.
Documentation and oversight
Technical documentation, event logging, training-data descriptions, human oversight and informing users they are interacting with AI. I build this together with the system rather than afterwards — retrofitting compliance onto a finished deployment costs several times more.
AI, GDPR and trade secrets
Assessing whether personal data and confidential information may be sent to a given model, and designing an architecture in which they never leave the organisation. Local deployment usually solves this at the root instead of papering over it with contract clauses.
Why me
Compliance advice usually comes either from lawyers without technical grounding or from engineers unfamiliar with the regulation. I work on both sides: I review international scientific publications and serve as an evaluating expert, while deploying models in production myself. I carried out the licence analysis, terms of service and AI Act / DGA compliance for my own service, gnss.day — this is not knowledge drawn solely from slide decks.
A running system, not a slide deck
I designed and launched a public scientific data service — from data acquisition, through an API backend with key management, to legal compliance and documentation in three languages. You can check it right now.
gnss.day — Earth Orientation Parameters via API
The service publishes Earth Orientation Parameters (EOP) — data used in precise satellite positioning, satellite collision avoidance and deep-space mission navigation.
- Automated source data retrieval and processing every 6 hours
- Backend with user accounts, API keys and daily rate limits
- Historical data back to 1973, interpolation to any epoch
- Legal compliance: source licence analysis, terms of service, AI Act, DGA
- Documentation and interface in three languages (EN / PL / DE)
GET /api/eop?date=2026-08-07
{
"mjd": 61259.0,
"x_pole": 0.1842,
"y_pole": 0.3271,
"ut1_utc": 0.0113,
"lod": 0.0009,
"source": "USNO"
}
I apply the same workflow to commercial projects: data → processing → API → documentation → legal compliance.
Why local rather than cloud
Sending company documents to an external model means they leave your infrastructure — which, for personal data, contracts or technical documentation, often conflicts with GDPR and client security policy. Locally hosted models solve this at the root: data never leaves the server, there are no per-query fees, and no dependency on a vendor changing its pricing. I run hardware (RTX A5000 48 GB and RTX PRO 5000 Blackwell 48 GB) capable of training and serving models without sending anything outside.
Bartłomiej Oszczak, PhD, Eng.
Researcher at the Department of Satellite Geodesy and Navigation, Faculty of Geoengineering, University of Warmia and Mazury in Olsztyn, Poland; also affiliated with the Polish Air Force University in Dęblin.
My field is positioning algorithms, GNSS satellite navigation and data processing — domains where a computational error has direct consequences. I bring that mindset to AI projects: correctness and measurable results matter more than an impressive demo.
Twice, as the only representative from Poland, I presented research results at the ION GNSS+ conference in the United States. I have collaborated with the Canadian Space Agency in Montreal and the European Space Agency.
I review papers for international scientific journals and serve as an evaluating expert. That work consists of verifying whether claimed results genuinely follow from the data presented — a habit that transfers well to assessing the promises made around AI.
- UWMDept. of Satellite Geodesy and Navigation, Faculty of Geoengineering
- PAFUPolish Air Force University, Dęblin
- Reviewerinternational scientific journals
- Expertevaluation of proposals and projects
- 2013ION GNSS+, USA — research presentation
- 2014ION GNSS+, USA — second presentation
- CSACanadian Space Agency, Montreal
- ESAEuropean Space Agency
What I work with
Hardware and tooling I have available — relevant because it determines what can be done without sending data outside.
AI where there is no data centre
Not every deployment needs the cloud and a permanent uplink. I build systems where the model runs directly on a device in the field — on a factory floor, inside a machine, on a drone. Data is processed on site; only the result goes out.
Small models on in-house hardware
I run and train language models on Jetson Orin Nano class devices — no cloud and no per-query subscription. I verify empirically what such a setup actually handles, instead of promising figures from a datasheet.
GNSS and precise positioning
Survey-grade and RTK receivers in my own lab: u-blox ZED-F9P (GPS, GLONASS, Galileo), NovAtel FlexPak6 OEM628 and FlexPak3. Differential corrections, NTRIP, position computation — over 100 GB of my own measurement data.
Prototypes and hardware integration
Raspberry Pi 4 and 5, StackyPi clusters, microcontrollers with HAT boards. From sensor, through data transmission, to the model making a decision — delivered as one working system, not a set of disconnected parts.
Let's talk about your process
The first conversation is free and without obligation. Tell me what consumes the most time in your company — I'll say plainly whether AI is the answer, or whether the money is better spent elsewhere.