Work with me
A senior engineer and AI architect for the parts of AI that have to hold up: governance, security and the engineering under them.
Radek Brus works with companies in regulated sectors as a contractor, as a consultant and on builds to order, with 14 years of engineering behind him, from embedded devices to enterprise AI platforms.
Three ways to work together
You work directly with the person who designs it and builds it.
Contract
On your team
A senior engineer and AI architect inside your team for a few months: architecture, hands-on build and reviews, on your stack and in your process.
For example: lead the architecture of an agent platform until your own team runs it.
Consulting
A fixed-scope review
One clear question, a fixed scope and a written result that your own people can act on.
For example: an Agent Handover Review before a vendor leaves, or a security assessment of an agent before its release.
Build to order
Built for you
You describe the outcome and the lab builds it: a prototype, an internal tool or an edge system, delivered with tests and documentation.
For example: a private assistant on NVIDIA edge hardware for data that must stay on site.
What you can ask for
Most of this work rests on tools the lab built and uses itself.
AI governance and agent management
An inventory of your AI agents: who owns each one, what it may touch and how it is stopped, with evidence that a risk or audit team can read.
You hold the model of your agent system, a passport for every agent, findings with their clauses and a signed handover pack.
Secure agent platform architecture
Agent identities, scoped tools over MCP and A2A, governed gateways and evaluation before release, on Microsoft Foundry or Google ADK.
You hold a reference architecture that your teams build on, with its decisions written down.
Security testing of AI agents
Red-teaming over REST, MCP and A2A as a regression test in CI, judged from what went over the wire.
You hold a repeatable test suite in your pipeline and a report that separates real findings from noise.
Calibrated evaluation
Decisions with an honest confidence value instead of one model acting as judge: local encoders first, the cloud only for uncertain cases.
You hold an evaluation that you can run again, with its calibration scripts.
Local and edge AI
Open-weight models on NVIDIA Jetson and DGX Spark, for data that must not leave the site.
You hold a running local service with an OpenAI-compatible interface and its benchmark scripts.
Off-grid radio and embedded systems
LoRa mesh radio, signed messages and firmware for field devices: the work behind Rifugio and Faro.
You hold tested firmware or a bench set-up, with its measurements.
How it starts
A first call
30 minutes on what the system does, what hardware and data it touches and what could cause harm.
A fixed-scope proposal
In writing: what will be done, what you hold at the end and what is out of scope.
The work
In your tools and your repositories, with regular check-ins and a written result at the end.
Practical
- Where
- Based in Locarno, Ticino. Remote across Switzerland and Europe.
- Working language
- English
- Sectors so far
- Medtech, pharma, industrial automation and building technology.
- Stack
- C/C++
- CUDA
- Python
- Go
- TypeScript
- C#/.NET
- ESP32-S3, LVGL, LoRa (MeshCore)
- llama.cpp
- NVIDIA Jetson Thor and DGX Spark
- Azure, GCP, AWS
- MCP
- A2A
- Certified
- Microsoft Multi-Agent AI Solutions Expert (AI-500)
- Azure Solutions Architect Expert
- Azure Security Engineer
- AWS Security Specialty
Tell me what you are building.
We start with a call of 30 minutes: what the system does, what it touches and what could go wrong. Then you get a fixed-scope proposal.