What do you need help with?

I provide AI engineering, machine learning, data science, automation, technical architecture, and research-oriented consulting.

My work ranges from early discovery and feasibility research through system architecture, implementation, deployment, optimization, and long-term support. I can work on complete systems or join an existing team to solve a specific technical problem.

I work on production AI systems, LLM applications, agentic systems, retrieval and search, multimodal AI, model training and fine-tuning, local and sovereign AI, scientific computing, data pipelines, analytics, and infrastructure supporting machine learning workloads.

I am particularly well suited to projects where AI intersects with difficult engineering, large or unusual datasets, privacy requirements, or multidisciplinary research.

Yes. I can help determine what should be built, design the architecture, and then implement it.

I generally prefer this combination because architectural decisions are much easier to make when the person designing the system also understands the engineering consequences.

Yes.

I can work from initial requirements and feasibility analysis through architecture, implementation, testing, deployment, performance optimization, documentation, and handover.

I can also productionize prototypes that already exist.

Yes.

A significant amount of engineering work involves understanding an existing system, identifying what is actually wrong with it, and improving it without needlessly rewriting everything.

I can audit architecture, reliability, performance, maintainability, model usage, data pipelines, retrieval systems, agent behavior, or infrastructure.

Yes.

Projects with unclear requirements usually begin with discovery. The purpose is to understand the problem, constraints, available data, business needs, and realistic technical options before deciding what should be built.

Initial discovery is free and carries no commitment. More substantial technical discovery and investigation is billable starting at €30/hour.

Yes, and I will explicitly recommend against AI when it does not provide meaningful value.

Sometimes a deterministic algorithm, database query, conventional automation, or process change is cheaper, faster, and more reliable than adding a model.

The objective is to solve the problem, not to insert AI into it for decorative purposes.

Projects where AI, automation, data, or advanced engineering can meaningfully improve efficiency, reduce repetitive manual work, accelerate analysis, or enable something that would otherwise be impractical.

I am also particularly interested in technically difficult or research-heavy problems, including multidisciplinary work involving fields such as medicine, biology, chemistry, physics, and scientific computing.

Generic software development that does not require AI, automation, data engineering, data analysis, or related technical expertise is usually not the best use of my time.

If discovery shows that I am substantially overqualified for what you need, I will say so. Where possible, I am also happy to recommend colleagues who can provide simpler implementation work at a lower rate.

I can integrate commercial LLM APIs when that genuinely solves the problem, but my work is not limited to wrapping an API in an interface.

My background includes model training and fine-tuning, model deployment, scientific computing, low-level systems programming, data engineering, analytics, ML infrastructure, privacy-preserving computation, sovereign AI, and production system architecture.

Sometimes an API call is the correct solution. Sometimes it very much is not.

Yes.

I design and implement systems involving tool use, planning, workflows, memory, context management, external services, permissions, human interaction, multiple models, and distributed execution.

This includes both conventional agent workflows and more complex custom orchestration systems.

Yes.

Where multiple specialized agents genuinely provide value, I can design communication, delegation, state management, permissions, lifecycle handling, and orchestration between them.

I do not introduce multiple agents merely because drawing more boxes makes an architecture diagram look advanced.

Yes.

That includes structured outputs, model routing, provider abstraction, streaming, tool calling, context management, reliability mechanisms, evaluation, observability interfaces, caching, and integration with existing systems.

Yes.

I generally prefer keeping model providers behind clean interfaces when practical so systems are not unnecessarily coupled to OpenAI, Anthropic, Google, or any other single vendor.

Different models can then be selected based on performance, cost, privacy, latency, or capability.

Yes.

I can evaluate the existing dependencies, determine what genuinely needs replacing, and migrate incrementally where possible rather than automatically proposing a full rewrite.

Yes.

I work with document ingestion, embeddings, semantic and hybrid retrieval, reranking, structured and unstructured data, citation and grounding systems, retrieval evaluation, and knowledge pipelines.

I can also investigate existing RAG systems that produce irrelevant, incomplete, or unreliable answers.

Yes.

The first step is usually determining whether the problem actually comes from the model, prompts, retrieval quality, context construction, tool design, data quality, state handling, or system architecture.

Changing prompts repeatedly and hoping the problem becomes frightened enough to leave is rarely an engineering strategy.

Yes.

I specialize in local and sovereign AI and can design systems that operate on infrastructure controlled by the client rather than depending entirely on external model APIs.

This can include local model inference, private retrieval systems, internal agents, and deployment on dedicated compute.

Yes.

Depending on the project, sensitive processing can remain entirely inside infrastructure controlled by the client.

I also work with privacy-preserving system design and confidential computing where stronger isolation requirements exist.

Yes.

I can evaluate model suitability, hardware requirements, quantization, inference performance, context requirements, deployment constraints, and the tradeoffs between local and hosted models.

Local inference is not automatically better, but it can be extremely valuable where privacy, cost predictability, latency, or independence matter.

Yes.

I can evaluate CPU, GPU, NPU, accelerator memory, system memory, inference engines, model size, throughput requirements, latency targets, and expected workloads.

The correct answer is not automatically "buy the largest NVIDIA GPU available."

Yes, provided the required models and supporting data can operate locally.

This can be useful for privacy-sensitive systems, unreliable network environments, embedded deployments, or environments where external services are prohibited.

Yes.

I work across machine learning, statistical analysis, data engineering, model evaluation, experimental design, training pipelines, inference systems, and production deployment.

Yes.

Whether training, fine-tuning, retrieval, prompting, or a conventional algorithm is appropriate depends on the available data, required behavior, cost, and expected improvement.

I will generally evaluate those tradeoffs before recommending expensive training work.

Yes.

That may involve examining the data, labels, preprocessing, model architecture, training procedure, evaluation methodology, distribution shifts, leakage, class imbalance, or whether the chosen metric actually corresponds to the real objective.

Yes.

I can design datasets, metrics, experiments, benchmarks, and evaluation pipelines for models or complete AI systems.

This is particularly important for LLM and agent systems, where anecdotal testing can make almost anything look impressive for twenty minutes.

Yes.

I am comfortable reading research literature, reproducing methods, translating mathematical descriptions into software, comparing competing approaches, and adapting research methods to practical systems.

Yes.

One of my strengths is being able to work directly with specialists outside traditional software engineering and understand enough of their field to communicate effectively and build useful technical solutions.

I am comfortable collaborating with experts in areas including medicine, medical imaging, biology, chemistry, physics, and other research-heavy disciplines.

Yes.

My research background includes multimodal medical diagnosis, and I work with computer vision, multimodal models, image processing, and systems combining multiple forms of data.

Yes.

This can include speech recognition, voice activity detection, speaker processing, real-time audio pipelines, low-latency systems, and local or privacy-preserving voice applications.

Yes.

Rust is particularly useful for infrastructure where reliability, concurrency, performance, predictable resource use, or low-level control matter.

I also work extensively with Python and use whichever language makes sense for the component being built.

Yes, but I would first determine whether doing so produces a meaningful benefit.

A rewrite may be justified by performance, concurrency, deployment requirements, reliability, memory use, or maintainability. Rewriting a perfectly adequate Python system merely because Rust has nicer error messages for one's suffering is less compelling.

Yes.

I work on modular, distributed, event-driven, asynchronous, and horizontally scalable systems, including architecture for stateful AI workloads, agents, workers, APIs, storage boundaries, and external integrations.

Yes.

Where practical, I prefer interfaces around infrastructure that is likely to change, including model providers, databases, storage, observability, transport, and external services.

This gives clients more freedom to change vendors or infrastructure later without rebuilding the entire application.

I strongly prefer using mature open-source libraries and existing tools where they already solve the problem well.

There is little value in spending client money recreating functionality that has already been implemented, tested, and maintained by a competent community.

I build custom components when existing solutions leave genuine architectural, performance, security, licensing, or functional gaps.

Yes.

The handling requirements depend on the project, but confidential, medical, or otherwise sensitive data should remain inside appropriate client-controlled infrastructure where required.

I will not casually download sensitive datasets onto personal systems simply because doing so is convenient.

Yes.

That can include explicit permissions, tool isolation, secret management, authorization boundaries, restricted data access, auditability, and safeguards around actions agents are permitted to perform.

Yes.

Prompt injection is primarily a system architecture and trust-boundary problem, not something solved by adding increasingly threatening sentences to a system prompt.

I can help design isolation, validation, authorization, provenance, tool permissions, and data-handling boundaries around model behavior.

Yes, when required by the project.

I can work within client-controlled development environments, systems, repositories, and infrastructure when security or data-governance requirements make that appropriate.

My main advantage is the combination of research ability, multidisciplinary understanding, system architecture, and hands-on implementation.

I learn unfamiliar technologies and scientific domains quickly and am comfortable working directly with subject-matter experts rather than requiring every problem to be translated into purely software terms first.

I also do not bill clients simply for the time required for me to learn a new technology or field. If I need to understand a new framework, protocol, scientific discipline, or technical concept before I can work effectively, that learning curve is my responsibility.

Research directly required to solve your specific problem, such as evaluating relevant state-of-the-art methods, experiments, or competing approaches, is normal project work and may be billable.

I do not see my role as simply receiving a specification and implementing it.

I want to understand the business, research problem, constraints, and reasons behind the requested system so that I can determine whether the proposed solution actually makes sense.

If I believe you are about to spend money solving the wrong problem, I will tell you, even if that means reducing the size of the project or losing the contract entirely.

Working independently means clients communicate directly with the person actually designing and implementing the system.

There is usually less organizational overhead, fewer communication layers, and more flexibility when requirements change.

Within an existing engagement I can generally adapt to new discoveries and requirements without turning every adjustment into another sales process, contract negotiation, or project-management ceremony.

Yes.

If discovery shows that the project does not require my expertise, I would rather tell you than charge a specialist rate for ordinary development work.

Where appropriate, I may also recommend colleagues who can perform simpler work at a lower rate.

Yes.

If an existing open-source solution solves the problem properly, using it is generally preferable to paying me to recreate it.

Custom software should exist because your requirements justify it, not because custom development creates more billable hours.

Rates depend on project size, complexity, technical requirements, and the nature of the work.

The hourly rate is agreed before work begins so you know what the engagement will cost.

They can.

Rates are determined by the complexity and requirements of the project rather than applying a single universal rate to every kind of work.

Yes.

Initial discovery is free and carries no commitment. This is intended to establish what you need, whether I can realistically help, and whether working together makes sense.

When discovery requires substantial technical work, research, investigation, architecture, data analysis, experimentation, or other deliverables, it becomes paid work.

Paid discovery starts at €30/hour.

No.

If I need to learn a technology, framework, protocol, or scientific field so that I can work effectively on your project, I do not bill that learning time.

Project-specific research that directly produces value, such as reviewing state-of-the-art methods, evaluating competing technical approaches, or investigating a research question, is billable.

No.

Work is normally invoiced monthly based on the hourly rate agreed before the engagement begins.

Invoices are due within seven days of being issued.

I can invoice in any practical currency, although EUR is preferred.

For other currencies, the invoice is calculated using the market exchange rate on the invoice date. I do not add a separate currency-conversion fee.

I am not currently VAT-registered in Slovakia.

For qualifying business clients elsewhere in the EU, the reverse-charge mechanism may apply. Clients outside the EU are generally not charged Slovak VAT.

The exact tax treatment is stated on the invoice as required for the particular transaction.

I generally prefer engagements of up to approximately 80 hours per month per client.

There is no fixed weekly allocation. Working hours can be adjusted around the client's timezone, team, and project requirements.

My normal preference is up to 80 hours per month, but I am open to exceptional full-time contract engagements where the project justifies it.

Yes.

For genuinely urgent work I can temporarily work up to approximately 12 hours per day.

Rush work is billed at 50% above the project's normal agreed hourly rate.

No.

I can work on a focused short engagement, a longer implementation, or an ongoing contract depending on what the project requires.

Yes.

Post-project support is available at €30/hour for simpler fixes and additions and €50/hour for more complex changes or additions.

I normally respond to email within one day.

For Slack or similar team communication, I aim to respond within 12 hours at most and usually respond considerably faster during normal availability.

For existing clients, I also try to remain reachable for genuine emergencies outside normal working hours when practical.

Yes.

I am happy to work under a client's NDA when appropriate. If the client does not already have one, I can also use OneNDA as a standard starting point.

This is defined by the agreement, but bespoke client work is usually assigned fully to the client.

If I integrate software, libraries, tools, or reusable components that I developed independently, I retain ownership of those underlying components while granting the client permanent rights to use and derive from them as agreed.

This prevents clients from becoming dependent on me while also allowing reusable technology to remain reusable.

Yes.

Proper code handover and relevant documentation are normal parts of engineering work and are not treated as some mysterious premium add-on.

The objective is that your team can understand, operate, maintain, and continue developing the system.

Yes.

I can work with private repositories, internal infrastructure, remote development systems, and client-managed environments according to the security requirements of the project.

Yes.

Most work can be performed remotely, and I can adapt working hours to collaborate with teams in different time zones.

Yes.

For on-site work, the client covers reasonable travel costs, travel time, and accommodation where accommodation is not provided directly.

I cover normal personal incidental expenses myself.

Yes.

I work remotely and can collaborate with companies and research teams internationally. Engagement structure, invoicing, and working hours can be adapted to the jurisdiction and timezone involved.