PAUL
JACKSON

A billing bug is not like other bugs. A broken button annoys a user; a broken checkout charges them. So when I owned QA on subscription flows, I treated every rupee, dollar and cent like it was my own money moving.

I ran 100+ REST endpoints through Postman — schema rules, auth headers, pagination, rate limits — and then went after the quiet part: webhooks. I replayed duplicate Stripe events to check idempotency, matched subscription state across Stripe, the backend database and the UI with SQL, and walked every upgrade, downgrade and cancellation path by hand.

The result was the most boring sentence in QA: zero billing sync issues shipped. Boring is the whole point. Nobody ever writes a support ticket that says "my invoice was exactly right," and that silence is the review I work for.

The content team's entire WordPress publishing flow was manual. Every post created by hand, every image downloaded and attached by hand, every upload scheduled by hand. It was honest work, and it was eating their week.

So I stole it back. I automated bulk post creation and image handling end to end — and the fiddly middle part nobody wants to do: linking each image to its post automatically, with a template sheet whose formulas matched file names to blog post names. Then schedulers for bulk chunk uploads, so a month of content could go out while the team slept.

Manual work dropped 60%. The best part: the content team still runs that flow today. The mark of good automation is that people forget it was ever manual.

The most dangerous AI output is not the one that looks broken. It is the one that reads beautifully and is wrong. A fluent lie passes every spell-checker and every tired reviewer's eyes — which is exactly why it needs a tester who reads like a skeptic.

Across weekly releases of AI content and discovery features, I probed for hallucinations, prompt injection edge cases, editorial rule violations, search index errors and broken category tags. I asked the questions a user would ask at 2am, then checked every answer against the source of truth instead of against my gut.

Testing probabilistic systems taught me the core skill of this whole discipline: deciding what "correct" means when the output is never byte-identical twice — and defending that definition at the release gate.