{
patrick = {
role = "Senior DevOps, AI and security engineer";
at = "SLTN, Azure Expert MSP";
builds = [ "identity" "delivery" "recovery" "AI infrastructure" ];
writes = [ "Terraform" "Bicep" "PowerShell" "Python" "Rust" "Nix" ];
believes = "every change reviewed, nobody over-privileged";
home = "https://patrick.kappen.io";
};
}I build the platforms other people deploy to, and the guard rails around them. Most of my work sits where infrastructure, identity and automation meet, and lately a good part of it is about letting AI agents in without handing them the keys.
- If it's not in Git, it doesn't exist. Portal clicks are drift with extra steps.
- A backup you haven't restored is a rumour. Recovery gets tested, or it isn't there.
- Standing admin access is a bug, not a convenience. Rights are granted just in time and expire on their own.
- AI agents get the intern treatment. A sandbox, a reviewer, and no production keys. I've watched an automated PR merge a feature, revert it, and still announce it in the release notes.
- Boring in production, interesting in design. The exciting part should be the pull request, not the incident.
At SLTN I own the platform code that our cloud teams and customer tenants run on, and I work with our CISO to turn security policy into automation instead of documents.
That currently means moving the whole organisation from Azure DevOps to GitHub and swapping every stored pipeline secret for OIDC on the way. Customer access runs through GDAP, Azure Lighthouse and PIM, all in Terraform, so changing who can do what is a pull request like any other. Around that sit automated landing zones and disaster recovery drills that fail over a full Azure Site Recovery plan, prove the machines come back, and always clean up after themselves.
Open models as coding agents. I pick, benchmark and tune open-weight models on vLLM, with my own suite for tool calling, long agent tasks, long context and many agents at once. Models are chosen on measured numbers for the actual workload, not on a leaderboard.
Agents with boundaries. Sandboxed execution, scoped credentials, human-approved Git promotion and agent setups declared in Nix, so every one of them can be rebuilt and audited.
A homelab run like production. Proxmox, Talos Kubernetes, GitOps, OpenTofu and Ansible; TrueNAS and Proxmox Backup Server with 3-2-1 backups on 10 Gbit; Prometheus, Grafana and Loki watching it all. My machines are a NixOS fleet with signed commits, built so that a compromised GitHub account can't push code onto them.
Graft turns a few lines of TOML into Podman Quadlet containers built from the Nix store. Describe what the container needs; Nix builds the rest.
- Studying for GitHub Actions and AI-103
- Building a blog on patrick.kappen.io
Last changed Oct 2026.
zerodawn-pi · pi extensions: bash jobs in a PTY, compact git, recipes, context pruning, tidy UI
- 🚀 v0.26.0 · Oct 2026 · release
bash: run a lone file view with read and a file write with write; run a lone file view with read and a file write with write. - 🚀 v0.25.0 · Oct 2026 · release
bash: convert git reads, file views and searches without refusal; bash: run each part of a chain with the tool of that part. - 🚀 v0.24.0 · Oct 2026 · release
let the agent start a job at once with background; bash: let the agent start a job at once with background.
zerodawn-nvim · Neovim addons in Lua, packaged as a Nix flake
- 🚀 v0.1.2 · Oct 2026 · release
block a tmux switch once, not once for each hook. - 🚀 v0.1.1 · Oct 2026 · release
start opens no second terminal when a client is paired.
zerodawn-tmux · tmux addons in shell, packaged as a Nix flake
- 🚀 v0.1.1 · Oct 2026 · release
start the agent as the window command, not as typed keys.
- Cloud: Azure
- Automation & delivery: Terraform · Bicep · GitHub Actions · PowerShell · Azure DevOps · Ansible
- NixOS & deployment: NixOS
- Linux: Ubuntu
- Containers & orchestration: Kubernetes · Podman
- AI: LiteLLM · Phoenix
- Datacenter & virtualisation: Proxmox
- Networking: Tailscale · Traefik
- Storage & backup: TrueNAS
- Monitoring: Grafana