The catalog audits itself

While I kept polishing Mi semana, I went back to a front that had stayed open since the beginning: even though the system already recognized products reasonably well, some cards were still the same product counted twice, and others claimed to be sold at a single store when it was actually available at several. For someone comparing prices, that isn’t a small detail — it’s the difference between trusting the number in front of them or not.

This time I didn’t settle for fixing cases by hand as I ran into them. I built a way to measure the whole problem: pull a random sample across the entire catalog, not just the cases I already suspected were wrong, and check for both kinds of error at once — things that should be together and got split apart, and things that got merged together that shouldn’t have been. The first measurement was uncomfortable: one in five cards was missing something. With that number as a baseline, every fix stopped being judged by whether it “looked better” and started being judged by how much it moved that number, up or down.

Up to that point there had also been a smaller but equally real problem: a brand name isn’t always clean data. The same brand shows up written differently at every store, some product lines belong to a bigger brand, some legal names aren’t the brand a shopper actually recognizes. I gave this its own process: each doubtful case gets resolved once, and that call stays on record, so it never has to be made again the next time it comes up.

With the catalog auditing itself instead of relying on anyone’s memory, it was time to put the house in order: the product’s code had grown fast, and before it kept growing it needed pieces with clear names and clear edges.