I have run orderly systems for a long time. Across all of it, one thing has been consistently true: the parts of my work I am most proud of are the ones where a messy, unpredictable operation became orderly, efficient, and measurably well-run.
That is not just a career observation. It is what makes me happy. A system that runs cleanly, a process that does not need babysitting, a control that fails loudly instead of silently: those fill me with a genuine sense of satisfaction and accomplishment.
Continuous improvement drives me. And I have learned that improvement without measurement is superstition, so ensuring we have the data to measure is critical to everything I build.
The thesis
The context that matters is cross-functional. An agent that knows one domain but not the others — alerts but not ownership, detections but not deployments — will mis-triage everything. Operations is downstream of operations in every other area of the house, so building effective agentic operations really means building the whole system well.
So I decided to prove it where I control the whole stack: at home.
The lab
My homelab is a production-grade k3s fleet: five Dell servers, GPU inference behind a gateway, GitOps delivery, a ClickHouse lakehouse joining DNS, network, LLM usage, and operations events, and an agent fleet with governance: admission gates, records, and graduation pipelines that measure whether autonomous work actually worked.
The point is not the hardware. The point is that the same concepts transfer across every domain I run: Kubernetes, DNS, home automation, a lawn mower, a blog. Consistency of pattern beats domain-specific heroics.
The principles
- Instrument first. No improvement without measurement; no claim without evidence.
- One fabric, many domains. Concepts transfer; the lakehouse and the patterns are domain-agnostic.
- Fail closed, prove the guard. Every control must be able to fail for the reason it claims.
- Durable over ephemeral. Nothing valuable lives somewhere that gets recycled.
- Governance as leverage. Gates and records make autonomy safe, not slow.
- Everything is a candidate lesson. Successes and failures both get captured and folded back.
- Humans direct, agents engineer. I own outcome and risk; agents own mechanism.
What this blog is now
This site is restarting. The old posts are gone; their author voice retired with them. What replaces them is the journey itself: patterns from the stack, written so you can use them, with the failures included, because the failures are the part everyone skips and the part that saves you the most time.
You will also get the honest runbook: what this costs in hardware, power, attention, and tokens. Almost nobody publishes that. It is the difference between a story and a tool.
One proof, right now
This very post was migrated with a verify-first, fail-loud, read-back discipline: a backup was verified before anything changed, every step was checked against the result it claimed, and nothing shipped unverified. That same discipline is what you will see applied to DNS outages, GPU contention, and mower automations in the posts to come.
Come back for the patterns. Bring your skepticism; the data will be here.