Open source
Open-source labs and tools
Hands-on courses, Ansible collections and tools for GPU, AI and Kubernetes infrastructure, free on GitHub.
On GitHub
Projects
Most are released under the Apache 2.0 licence. Every version is pinned and every build is tested end to end. Independent community work, not affiliated with NVIDIA, Docker or Red Hat.
Hands-on labs
Run:ai OSS Hands-On
Open-source NVIDIA Run:ai 3.0 on one GPU workstation, built as training material: an Ansible install on RKE2 VMs with the GPU passed through, or on k3d in 15 minutes, then 13 labs on GPU fractions, sharing, quotas, gang scheduling and more, each with a check that fails if the feature misbehaves.
github.com/omerfsen/runai-oss · 13 videos on YouTube · the article
OpenShell Hands-On
A self-paced course on NVIDIA OpenShell, the runtime that lets AI agents work without open access to your data, secrets or network. 20 labs from a fresh Ubuntu machine to a real coding agent under a policy you wrote and formally verified; every lab has automatic checks and a recorded solution.
github.com/omerfsen/openshell · 19 videos on YouTube · the article
Docker Sandboxes Hands-On
The short twin of the OpenShell course, for Docker Sandboxes (sbx): coding agents such as Claude Code run in microVMs with their own kernel, and a proxy decides every outbound connection. 9 labs, about 2-3 hours, each with automatic checks and a recorded solution.
github.com/omerfsen/docker-sandbox · sbx vs OpenShell
NVIDIA Infra Controller lab
NICo on a single KVM host, as an Ansible collection: a three-node site cluster, NICo built from source, and virtual “bare-metal” servers with mock Redfish BMCs that NICo discovers, ingests and installs an OS on, one of them with a GPU passed through.
github.com/omerfsen/nvidia-nico
OpenShift lab on KVM
A current OpenShift 4.22 cluster on one libvirt host, built by Terraform: the machines, the network, the DNS and load balancer the installer will not create, and the install itself. About 25 minutes from apply to a working console, and terraform destroy takes it all back out.
github.com/omerfsen/openshift-lab
Ansible collections for an AI factory
They chain together: nvkr builds the artifacts, nvke boots the cluster from them, aicr puts the GPU software stack on top.
omerfsen.nvkr
Builds NVIDIA Kubernetes Engine artifacts with NVIDIA Kubernetes Recipes on one host, checks every digest and hands them to omerfsen.nvke.
github.com/omerfsen/nvidia-nvkr
omerfsen.nvke
The NVKE management cluster: reference node image, TPM-attested nodes and a BGP-announced API VIP, optionally all on KVM with a GPU passed through.
github.com/omerfsen/nvke
omerfsen.aicr
NVIDIA AI Cluster Runtime recipes (GPU Operator, DRA driver, NVSentinel, KAI Scheduler, monitoring) applied with helmfile or Argo CD and validated against the recipe.
github.com/omerfsen/aicr
Tools
NVSentinel MCP server
An MCP server in Go for testing NVIDIA NVSentinel: an AI client reads what it reports, injects a fault into a simulated GPU and checks that the node is cordoned and drained. Published to the MCP Registry, with release binaries for Linux, macOS and Windows.
github.com/omerfsen/nvidia-nvsentinel-mcp
Let’s work together on your next project
Tell us about your environment and goals. We’ll come back with a clear, practical plan.