How I ship AI workflows that survive contact with operations

Most AI demos end at the applause. Operations begin the next morning, when inventory is wrong, a product image failed to generate, and three people are waiting on a workflow that “worked in staging.”

The pattern is familiar: a promising agent, a polished interface, a weekend prototype. Then the same system meets merchant catalogs, rate limits, partial failures, and humans who will not re-run a broken job “just to see.” Production is not a harder demo. It is a different problem.

Start from the work, not the model

Before choosing models or frameworks, I map the operational path: who triggers the job, which systems must agree, what “done” looks like, and what happens when step four of seven fails. That map usually points to boring infrastructure first — queues, credentials, backups, least-privilege access — and only later to the clever part of the stack.

If the runbook is unclear, the model will not save you. It will only fail faster.

On Travel Avenue and the wider Osaka Group stack, that means n8n pipelines wired to catalogs and media, internal tools behind WorkOS, Docker Compose on AWS, and monitoring that pages people before customers notice. The “AI” layer sits on top of habits that operations already trusts.

Own the URL

Each post on this report is a full page with its own address — readable by people and by search engines. The homepage links here; it does not host the essays inline. That keeps the portfolio scannable and gives long-form writing room to rank and age well.

This is sample copy for the first live article template. Replace with the final essay when you publish.