Autonomous Infrastructure Logs

Before the architecture diagrams and the agent logs, here's the actual reason this homelab exists — and what this series is going to cover.

Running a homelab from Central Asia means CGNAT, no static IP, and a power grid that isn't always polite. Here's the hybrid architecture that makes none of that matter.

This is how the Helsinki edge terminates TLS, issues certificates for services that were never publicly reachable, and routes traffic home over an encrypted tunnel without opening a single port on the local firewall.

LLMs are non-deterministic. Infrastructure-as-code has to be deterministic. Here's the prompting discipline and CI gating that lets Claude and Gemini write real Terraform and Ansible without that being a contradiction.

Here's the move from static alerts to a small catalog of AI agents that diagnose cluster problems in read-only mode, propose a fix, and — inside a narrow, pre-approved scope — apply it before I've had coffee.

If-this-then-that automations don't share a model of the house. This is an in-progress log of building one that does — starting with a floor plan, a taxonomy, and the first Shelly sensors about to feed it real data.

Here's a working framework for when to adopt existing open-source software and when it's actually cheaper, long-term, to have Claude and Gemini help you build the smaller, sharper tool instead.

The software-defined stack we've built so far still sits on a flat physical network. Here's how we tore that down into a zero-trust 802.1Q VLAN topology with pfSense enforcing least-privilege at every boundary.

Handing an LLM an open-ended task is a recipe for hallucinated, undocumented spaghetti. Here's how I applied strict Agile discipline — Tracks, a Technical Manager Agent, and a non-negotiable human review gate — to turn AI-written code into a predictable delivery pipeline.