Vellox: alerts people can trust

01

How it works

Every morning a scheduled job reads each competitor’s pricing page, sitemap, job board, blog and homepage, using a real browser where prices load through JavaScript. Each read is compared with the last one. Plain code decides what changed, and only then does an AI model judge whether it matters and describe it.

02

Why rules come first

A model asked to spot changes can invent them. Here it only sees changes the code has already found, along with the evidence, and returns structured output that is checked before use. If the AI is unavailable, only clear-cut factual changes get through, and only a real price change can trigger an instant alert.

03

Telling a real change from a bad load

• Prices vanishing usually means the page failed to load. Vellox sets that check aside. Only after three in a row, with the page still recognisably the same, does it report “took the prices off their pricing page”.
• A bad first snapshot used to make prices look new the next day, sending a false alert to a brand-new user. The checker now starts a fresh baseline instead.
• Plans reordered by an A/B test are compared as sets, so they don’t count as a price change.

04

Keeping it running

Trial reminders were skipped when a job started late, because the logic rounded days left. A simulation of 1,920 users with realistic delays found the old rule missed a reminder for 174 of them, and the new one for none. After failed Vercel builds twice left the live site on an old version for hours without notice, I added CI with type checks, lint, tests and a production build. The project now has about 140 automated test cases.

05

Reaching the region

It has an Arabic landing page and payment in Egyptian pounds through InstaPay or mobile wallets.

06

What isn’t solved

Careers pages that show no roles still fail rather than resolve, because auto-accepting them would cause false hiring alerts if the page recovered. Vellox launched in October, and it’s too early to claim traction.