opsira

Running AI automation in production

Plenty of people write about building automations and AI agents. This is about running them: what breaks at month three, how they fail without telling you, and what it takes to keep them alongside real orders, real customers and real money. Everything here comes from systems in production, not from a demo built to illustrate a point.

Start here

7 guides. Each covers one area end to end and hands off to the notes underneath it.

n8n · 3 notes

Running n8n in production

What changes when a workflow tool stops being a toy: versioning, webhook reliability, exposure, and the failures that produce no error at all.

Support · 3 notes

Running an AI support desk

Inbound mail into tickets, a model that classifies rather than replies, and the traps that turn a support desk into a machine talking to machines.

Orders · 6 notes

Automating orders, payments and returns

The unglamorous half of an automated business: matching money to records, references humans can read, and getting your hardware back.

Visitors · 7 notes

Knowing where your traffic actually comes from

Per-visit records instead of aggregates, first touch instead of last click, and why a working channel can read as zero.

Being found · 11 notes

Being cited by AI search

What actually moves AI citations, what is measurement noise, and the boring technical failures that stop you being read at all.

Email and DNS · 8 notes

Email that actually arrives

SPF limits, sending domains with no reputation, tenant moves that lose mail, and the failures that report success.

Infrastructure · 19 notes

Self-hosting it reliably

Deploys that did not deploy, firewalls that are not the firewall, growth that crosses a limit, and backups nobody has restored.

All 57 notes

Why this exists

Most writing about automation is produced by people who built a demo. It is confident, tidy, and wrong in exactly the places that only show up once something is carrying real traffic.

The failures worth documenting are rarely the loud ones. A change that throws an error costs ten minutes. A change that applies cleanly, reads back correctly, and quietly does nothing can cost days before anyone notices the thing they fixed is still broken. Those are the ones written up here.

Each note is narrow on purpose. One problem, the mechanism behind it, how to confirm you have it, and what actually resolves it. No introductions, no filler, and no pretending a workaround is a fix.

Need help with any of this?

These notes are free and always will be. If you would rather someone just set it up, or you are stuck on something similar, get in touch at hello@opsira.io.