This site runs itself

The page you're reading isn't on a big cloud provider. It's served from a container on hardware I own, behind a reverse proxy I run, on a fleet that largely operates itself. That last part is the interesting bit.

The principle: deterministic first

The core rule of the whole system is that steady-state operations should be boring. Monitoring, health checks, restarts, verification — all of that is plain deterministic code that runs the same way every time. No AI in the steady state, no cleverness, no surprises.

Every service in the fleet gets probed on a schedule. When a probe fails, the system doesn't page a human — it follows a playbook: restart the service, verify it actually came back, log the incident, close it. The loop is designed to close routine failures without human paging.

Escalation is where judgment lives

Some failures don't fit the playbook. A service that keeps crashing after a restart, a disk filling faster than it should, something that looks wrong but isn't in the runbook. Those cases escalate — first to AI agents that investigate and propose a fix, and past a certain risk threshold, to me.

The important design decision: anything risky waits for a human yes. Automation gets full authority over reversible, well-understood actions and zero authority over destructive ones. That boundary is enforced by the system, not by good intentions.

Why bother?

Operations time is the tax on everything else. Every hour spent babysitting servers is an hour not spent building. The fleet runs a few dozen services — the ventures on the front page, their databases, and the tooling that watches all of it — with routine operational work handled automatically.

The live status chip on the homepage is part of this: it's a real server-side check of the public venture sites, not a decoration. If one of them goes down, that chip goes amber — and somewhere behind it, a playbook is already running.

What self-hosting costs

Self-hosting trades recurring platform dependence for upfront design, hardware, and ownership of failure handling. The automation had to be built, tested, and hardened through real iterations before I trusted it overnight.

The compounding return is practical: a workflow solved once can be reused, and every new service inherits the operating model from day one. This site did — it was monitored before it had content.