Agentic AI · on your own infrastructure
AI training for businesses
Role-based courses to build, run, secure and govern AI agents on your own servers: from testing with Ollama to production with vLLM, Kubernetes and MCP, on Ubuntu or Rocky Linux with support from Canonical or CIQ.
We are a Canonical and CIQ partner.
the machine, bottom up ↑
Path for AI infrastructure
Agentic AI training by role
Each path lists the courses in order, with the days it adds up to and where to start if your team already has the basics. Courses with a course page are delivered at your offices or remotely, with a live instructor; those marked “tailored” are prepared on your own environment.
AI infrastructure
- Ubuntu system administration basics UBU1244 days
- Advanced Ubuntu system administration UBU1344 days
- Bare-metal servers with MAAS UBU4473 days
- Local LLMs and GPU infrastructure SXIA033 days
- Ceph: deployment and administration3 days
- Shared GPUs on Kubernetes: GPU Operator, DRA and Kueue SXIA122 days
The path adds up to 19 days. If you already run Linux, start with MAAS: 9 days.
AI platform
- Kubernetes with MicroK8s UBU1803 days
- MLOps with Ubuntu UBU6103 days
- From Ollama to production: vLLM and llm-d on Kubernetes SXIA083 days
The path adds up to 9 days. If you already use Kubernetes, start with MLOps: 6 days.
AI engineer: agents
- LangChain: RAG and agents on-premise SXIA052 days
- AI agents for businesses SXIA013 days
- MCP servers: connecting agents to your systems SXIA092 days
The path adds up to 7 days. If you have already built a RAG system, start with agents: 5 days.
NextRAG with Docling
AI security
- AI security: OWASP LLM and red teaming SXIA072 days
- Agentic AI security: permissions, MCP and isolation SXIA112 days
The path adds up to 4 days. If the team does not administer Linux, Advanced security on Ubuntu comes first: 3 days.
AI governance
The path adds up to 2 days.
Other profiles
- Ansible, AWX and Juju (DevOps) UBU2945 days
- Docker, Podman and Kubernetes (DevOps)3 days
- Big data and Apache Spark for AI (data) SXIA063 days
- Deep learning with PyTorch and TensorFlow (data) SXIA043 days
- Copilot and ChatGPT for businesses (business) SXIA022 days
- AI architecture workshop on your own infrastructure (IT management)tailored
This is not a path: each team picks the course it needs.
NextWhat an AI factory is and how to build one with open source
The infrastructure for agentic AI, from GPU to agent
Everything is upstream software. The system and the platform are commercially supported by Canonical on the Ubuntu stack or by CIQ on the Rocky Linux stack, and we are a partner of both, so the course and the installation are handled by the same team.
| Layer | With Canonical | With CIQ |
|---|---|---|
| GPU servers | Lenovo or Supermicro, with either stack | |
| Operating system | Ubuntu, with CUDA in the Ubuntu 26.04 LTS repository | Rocky Linux RLC Pro AI, with validated CUDA, DOCA-OFED and PyTorch |
| Provisioning | MAAS | Warewulf |
| Cluster | MicroK8s · Kubernetes | Slurm · Apptainer |
| Storage | Charmed Ceph | Ceph, supported by us |
| Models | Ollama for testing, vLLM and llm-d for serving · Charmed Kubeflow and Charmed MLflow if you need MLOps | Ollama, included in RLC Pro AI · vLLM for serving |
| Agents | LangGraph · MCP, with either stack | |
More training
Official Canonical courses
Ubuntu administration, Kubernetes, MAAS and MLOps, with the official Canonical syllabus.
Canonical catalogueAutomation with n8n
n8n courses to automate processes with agents, and the implementation if you would rather we did it.
n8n introductory coursen8n advanced courseAI automationOfficial IBM training
watsonx, Watson Studio, Discovery and Assistant, with the official IBM course codes.
IBM AI coursesvLLM on IBM Power, without a GPUFrequently asked questions
What is agentic AI?
AI systems that do more than answer: they plan steps and use tools, such as databases, APIs or applications, to complete a task. In a business the usual approach is to give them limited permissions and require human approval before sensitive actions.
Do we need GPUs to run agentic AI on our own servers?
To serve models to several users at once, in practice yes: the GPU determines how many requests the server handles and how quickly. For testing with small models a powerful CPU is enough. In the local LLM course we measure against your workload before you buy hardware.
What is the difference between Ollama and vLLM?
Ollama is the simplest way to try a model on a laptop or a server: it installs in minutes and works well for a few users. vLLM is built to serve the model to many people at once, because it handles many requests in parallel and makes better use of GPU memory. The usual route is to test with Ollama and move to vLLM as usage grows; if one server is not enough, llm-d spreads the load across several.
Ubuntu or Rocky Linux?
Both work. Ubuntu, supported by Canonical, has Kubeflow and MLflow packaged by Canonical (Charmed Kubeflow and Charmed MLflow) and CUDA in the Ubuntu 26.04 LTS repository. Rocky Linux, supported by CIQ, fits if your servers already belong to the Red Hat family, and its RLC Pro AI edition comes with validated CUDA and PyTorch. We choose the path based on what you already have.
What if our servers are IBM Power?
Yes. vLLM runs on IBM Power, even without a GPU, and IBM publishes an official MCP server for IBM i that gives agents access to Db2 for i with each user’s permissions. We are an IBM Business Partner and adapt the paths to Power.
Can the courses be run on our own servers?
Yes. They are delivered at your offices or remotely, with a live instructor, and tailored courses are prepared on your own environment: your GPUs, your system and your test data.
What if we would rather you built it for us?
We do that too: GPU servers, system, cluster and models, with a single point of contact. AI inference on your own servers →
Request a training proposal
Tell us your team’s role, how many people it has and what infrastructure you use. We will reply with a proposal: which courses, in what order and where they are delivered.
Message us on WhatsApp
