We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Run language models on your own hardware, so your team and its agents can use AI on data that must not leave the building. Leave knowing what to buy and what it costs to run.
Rebuilt on your team's work
Run privately for your team, on dates and at a length that suit you.
Every module ends with something participants can repeat on Monday morning in their own work and tools, so the team gets more done without adding people.
Judge whether self-hosting solves your problem, or whether you would be buying hardware to get round a contract clause.
Pull, run and prompt an open-weight model on your own machine inside the first morning.
Work out what memory, GPU and disk a given model needs before anybody orders anything.
Put a model behind an API that several applications can call at once without it falling over.
Trace where your prompts and documents go, and prove they stay on your network.
Put hardware, power and your own maintenance time next to an API bill and compare them properly.
Most of the time goes on hands-on exercises. Module order and depth are set for your group before the first day.
Every length uses the same modules. A longer booking takes the team further through the same curriculum.
1 day
A day to get a model running on your own machine and to know which one it ought to be. The version for a team working out whether any of this is realistic.
2 days
Adds serving the whole team from one machine, and running it day to day: access, updates, monitoring and cost.
We settle the length when we talk about your team, and it can still change afterwards.
Your version starts from the hardware you have or plan to buy and the documents your team needs to work with. Models are compared on your own tasks, and serving is set up the way your infrastructure already runs, from a single GPU server to an existing cluster.
How a booking worksIn your office, with the team in the room. Best for a first rollout.
Live over video, in shorter blocks across several days.
A day on site to start, then shorter remote sessions on real work.
Every course runs privately for one team, on dates we agree together.
Every module can be rebuilt on your team's own work and tools. Tell me what the team works on and I will tell you what I would change.
AI agent trainer and software engineer
I build LLM systems that run in production, including retrieval with Qdrant and pgvector, so the course measures models against work they have to do rather than against benchmarks.
From engineering teams I trained to work with AI agents.
I worked closely with Luka as the topic sponsor for a workshop on integrating AI meaningfully into everyday development work.
He took the time to understand our participants' needs and adapted his approach to the tools and working environments available.
Participant feedback reinforced my own positive experience, with especially strong recognition of his subject expertise and the balance between theory and practice.
I wholeheartedly recommend Luka as a trainer to any team looking to build practical AI skills.
What teams ask before they book.
No. This is an infrastructure course, and nobody trains a model. It assumes you are comfortable with a Linux command line and with containers, because that is where the day is spent.
Not for the first day. Small models run acceptably on a laptop with 16 GB of memory, which shows that useful self-hosting starts smaller than most teams assume. For the serving material, bring a machine with a suitable GPU if you have one; otherwise we use smaller models that run on the laptops.
The current open-weight families, chosen fresh before each course, because any fixed list goes out of date within months. The module teaches the method for choosing one, which stays the same.
It solves one part of it and leaves other parts untouched, and working out which is a scheduled part of the day. Running a model locally means your prompts stay put. It does not answer who may query it, how long the logs live, or what happens to what it produces. Bring the question your legal team asked.
For some tasks plainly yes, for others plainly not, and the course is blunt about where the line runs. Classification, extraction and drafting against your own documents work well. Anything needing frontier reasoning still does not. You leave able to place your own use cases on that line yourself.
Often. A hosted model is stronger and cheaper to start with, and an EU data processing agreement settles many data protection questions. Self-hosting pays off when the data may not leave the building at all, when usage is high and steady, or when you need a model that does not change under you. The course settles that question before anyone installs anything.
Have a question about your team's setup?
Tell me what your team works on and how it uses AI today. A few lines are enough. I reply myself, usually within a day, and if this course is the wrong one I will point you to the one that fits.
In English or German, on site, remote or hybrid.
Self-Hosted LLM Training
A few lines are enough
Engineering teams and providers I have delivered AI training for