AI ENGINEERING · PRIVATE INFERENCE · CONTROLLED OPERATIONS
Safematix engineers AI systems for constrained and high-security environments: local and private inference, retrieval over your own data, training and tuning, quantization, and agent enablement. The work is engineering and enablement—building something that runs inside your perimeter and can be operated—not a claim to train frontier models or to guarantee what a model will answer.
AT A GLANCE
01 / BUYER FIT
Many AI initiatives start with a model and end with a demo. Safematix starts from the operating constraint: which data may be processed, where the system must run, who operates it and how its behavior can be checked. That usually changes the architecture before any model is chosen.
01
Your material is regulated, contractually restricted or simply not permitted to leave your infrastructure, so a hosted API is not an option. Safematix builds the inference and retrieval stack inside the environment you control, and documents the network path each request takes so the boundary can be reviewed rather than assumed.
02
You have a use case with an owner, an approval process and an expectation that the result will keep running. Safematix delivers a working service with deployment, evaluation, logging and an operating guide, so the result can be reviewed, handed over and maintained instead of remaining a notebook on someone’s laptop.
03
A model that works in a demo may not fit the GPU, memory or latency budget you actually have. Safematix measures where the cost sits, then applies quantization, batching and serving changes and states plainly what each trade-off costs in quality and throughput, so the decision is documented rather than guessed.
04
Your team will own the system after handover. Safematix works alongside them on the actual codebase, explains the parts that are easy to get wrong, and leaves runbooks and evaluation harnesses so the team can change a model, a prompt or a retrieval index without losing the ability to tell whether it still works.
02 / CAPABILITIES
An engagement selects from the work below and fixes its output in writing. Scope is honest about limits: Safematix engineers, tunes and operates systems on the hardware and data you can provide. It does not claim to pre-train frontier models, and no engagement can promise that a model will always produce a correct answer.
01
Safematix deploys and operates model serving on infrastructure you control—on-premises, in your cloud tenancy or on dedicated hardware. Work covers runtime selection, GPU and memory sizing, batching, caching, health checks and the network path, so inference runs inside the perimeter with a documented operating model rather than a vendor dependency.
02
Safematix builds retrieval pipelines over the documents and records you are permitted to use: ingestion, chunking, embeddings, index maintenance and access filtering. The result cites the passages it used, respects the permissions of the requesting user and can be re-indexed when source material changes—the parts that decide whether a RAG system is usable in production.
03
Where a general model is not sufficient, Safematix prepares datasets, runs fine-tuning or continued training within the available compute, and builds an evaluation set before and after. The output reports measured behavior on that set with its limits and known failure modes; it does not present a benchmark score as a guarantee of correctness in your domain.
04
Safematix profiles where latency, memory and cost actually sit, then applies quantization, distillation, batching or serving changes and measures the effect on quality and throughput. Each change is reported with the trade-off it introduced, so your team can decide which configuration fits the operating budget instead of accepting a single unexplained number.
05
Safematix connects models to tools, data and internal services with explicit boundaries: which tool a model may call, with whose authority, with what audit record and what happens when a call fails. This is the engineering counterpart to AI security assessment—build the integration so the access decisions are visible and reviewable.
03 / DELIVERABLES
The written scope selects the relevant deliverables below. A focused performance engagement does not automatically include a full retrieval pipeline or an agent integration, and a system handover is not an ongoing operations contract unless that is separately agreed.
01
A description of the system as built: components, data flows, the perimeter each request crosses, the hardware and the operating assumptions. It records the constraints that shaped the design—data residency, latency, budget, available compute—so later readers can tell which choices were deliberate.
02
The working inference, retrieval or agent service in the agreed environment, with its configuration in version control and reproducible deployment steps. Safematix confirms it starts, serves and shuts down cleanly in that environment before it is considered delivered.
03
A task-relevant evaluation set with the results of the delivered configuration, including the cases where it performs poorly. The record states what was measured, on which data and under which limits, so a later change can be compared against a known baseline rather than an impression.
04
How to start, monitor, update and roll back the service, including the signals that indicate degradation and the steps for replacing a model or re-indexing data. It names the operational responsibilities that stay with your team and the assumptions that need re-checking when the environment changes.
05
A working session with the engineers who will own the system, covering the parts that are easy to get wrong: prompt and index changes, evaluation drift, access filtering and cost behavior. The goal is a team that can change the system and still tell whether it works, not a dependency on the original author.
04 / EXPERIENCE
You work directly with Robert Ressl. His public portfolio documents model artifacts, inference engineering and AI-security work; these are descriptions of his engineering background, not invented customer results or claims about a particular model’s accuracy.
Robert’s public portfolio documents operating a Kubernetes-based inference platform and publishing model artifacts. That is hands-on experience with serving, GPU scheduling, rollout and the operational problems that appear after a demo works—the part of an AI system that determines whether it can stay in production.
His public work includes training and tuning models and publishing the resulting artifacts and engineering notes. It covers the practical parts: dataset preparation, compute constraints, evaluation and the difference between a result that looks good on a benchmark and behavior that holds on real input.
Robert’s public AI-security work includes MCP tooling and controls around model-driven tool use. Building and assessing these systems are related skills: the same understanding of data access, tool authority and audit records that hardens an agent is what makes an integration operable in a regulated environment.
05 / Approach
The work starts from what the environment allows, not from a model name. Each step makes clear what Safematix builds, what it measures and what your organization must provide or decide.
Safematix clarifies what the system must do, which data it may use, where it must run and what latency, budget and hardware are available. The output is a short architecture proposal with the assumptions stated, so the first decision is made before compute is spent.
Your responsibility Name the use case owner, describe the data and environment at a high level and confirm the available hardware and budget.
Safematix builds an end-to-end path—data in, model, output, and the retrieval or tools it needs—in the agreed environment. The first version is deliberately small, so the team can see the whole system work before optimization and can confirm the interface and operating model.
Your responsibility Provide access to the agreed environment, sample data and the people who will operate the system.
Safematix profiles latency, memory and quality, applies quantization, batching or retrieval changes and measures what each change cost. Improvements are reported with their trade-offs, so the delivered configuration is a documented decision rather than a series of experiments left in place.
Your responsibility Confirm the operating budget and quality threshold, and decide which trade-offs are acceptable.
Safematix hands the system to the engineers who will own it, with a runbook, the evaluation set and a working session on the parts that are easy to break. Where agreed, Safematix supports the first model, prompt or index change so the team practices the update path while support is still available.
Your responsibility Assign the owning engineers, take over operations and confirm which ongoing responsibilities remain with Safematix, if any.
06 / COMMERCIAL TERMS
AI engineering work is quoted against the agreed task, environment, hardware and deliverables rather than sold as a fixed package. Safematix confirms objectives, inputs, deliverables, responsibilities, verification, fees and the start date in writing. The published day rates provide the commercial starting point; the applicable rate and effort are confirmed in the offer.
Flexible engagements: CHF 2’200 per working day. Longer, clearly scoped remote projects: CHF 2’000–2’100. Short-notice OT, red-team, and high-security engagements: CHF 2’400–2’600, subject to scope. Package prices are fixed-scope equivalents at CHF 1’850–2’100 per working day; the day rate applies when the scope cannot be fixed in advance.
All package prices and day rates exclude applicable VAT and pre-approved travel and third-party costs, unless expressly included in the order form. Fixed-price packages apply within their stated scope. Scope, authorization, and start date are confirmed in writing before work begins. Regulatory mapping is technical, not legal advice. Savings and compliance outcomes are not guaranteed.
Discuss an AI system07 / FAQ
No, and the engagement will not claim to. Safematix works with existing open models: it fine-tunes, tunes, quantizes, serves and integrates them within the compute you have. Pre-training a frontier model requires resources and data far beyond a typical engagement, and a proposal that promises it should be examined closely.
No system built on a language model can be guaranteed to answer correctly every time, and Safematix will not represent one as if it could. What the engagement delivers is measured behavior on a task-relevant evaluation set, the known failure modes, and controls—retrieval citations, refusal paths, human review where needed—so your organization can decide what the system may be trusted to do.
Not necessarily. Local and private inference is a core part of the work: the model, retrieval index and logs run on infrastructure you control. Where a hosted service is acceptable for part of the task, that is a decision recorded in the architecture, with the data path documented so it can be reviewed against your obligations.
It depends on the model size, the context length and the latency you need. Safematix sizes the requirement during scoping and states what fits on the hardware you have, what would require more, and which quantization or serving choices change the answer. No specific GPU vendor or model is assumed in advance.
AI security assessment asks what an attacker or a misused model can reach; AI engineering builds the system and its controls. They share an understanding of data access, tool authority and audit records, and can be combined—but one does not automatically include the other. Safematix will state which is being commissioned in writing.
Describe the task the system should perform, which data it involves at a high level, where it must run and what hardware or budget you have. You do not need a finished specification to begin. Please keep credentials, personal data and sensitive architecture details out of the initial form; secure sharing can be agreed during scoping. I reply personally, usually within one business day (Mon–Fri, CET).
08 / Contact
Tell Safematix what the system must do, which data it may use and what constrains it. Safematix can then propose an architecture, a scope and an evaluation approach. Keep credentials, personal data and sensitive architecture details out of the first message.
Discuss an AI system