Why I'm Doing This: An Introduction to the Homelab

I don't need a hybrid cloud, a Proxmox cluster, or a fleet of AI agents to watch a media server. None of it produces anything I couldn't get from a single box and a cron job. So it's worth saying plainly, before the first architecture diagram shows up: the lab was never the deliverable. It's the excuse.

What I'm actually testing

Production systems punish experimentation. A homelab doesn't. If an AI agent misreads a network log at 3am and takes down a media server for an hour, that's a Tuesday. If it happened on infrastructure that mattered, that's an incident review. So this is where I get to ask the question I actually care about, with real consequences small enough to survive being wrong: how far can you push AI agents into real infrastructure work — provisioning it, watching it, repairing it — before their judgment quietly stops being good enough?

That's a more interesting question than "can an AI run a homelab." The homelab part is mostly solved already. What isn't solved is where the judgment breaks down, and whether giving agents a narrower, more accountable role — instead of one generalist doing everything — pushes that breaking point further out.

The shape of the project

Two constraints made the shape of this lab non-negotiable from the start. First, geography: running physical infrastructure out of Kazakhstan means none of the usual assumptions about static IPs or simple port-forwarding hold, so the network design had to solve that before anything else could get built on top of it. Second, everything past the network layer — provisioning, monitoring, remediation, even home automation — is a candidate for handing to an agent, provided the agent's role is scoped narrowly enough that I can trust the boundary, not just the individual decision.

What this series covers

This series follows the lab roughly in the order it got built:

The hybrid architecture itself — a cloud gateway and a home cluster that don't share a network in any conventional sense, and why that's a feature.
The routing and security layer that makes the public edge trustworthy enough to front a private lab.
Handing Terraform and Ansible to Claude and Gemini, and the discipline that keeps non-deterministic code generation from touching a live cluster unreviewed.
AgentOps — the AI agent running in an actual SRE role across the lab, and why that's a different goal than smarter alert scripts.
A cyber-physical digital twin for the house itself, still in progress, written honestly as a build log rather than a finished result.
The build-vs-buy framework this whole project keeps forcing me to answer — when a custom tool is worth owning, and when it isn't.

More topics will get added to the list as the lab grows — this isn't a fixed-length series.

Why publish any of it: a fix I can't explain in writing is usually a fix I don't actually understand, and the failures here are more useful to write down than the successes ever are.

Next post: the hybrid architecture — a homelab in Kazakhstan, fronted by a cloud gateway in Helsinki.