A faster judge for the catalog
Judging whether two products are the same, case by case, doesn't scale. I brought in an outside service that clears the obvious cases on its own, saving the costly judgment for what disagrees.
The work of recognizing when two products are the same — the problem I described a few weeks ago as the hardest one on this project — never really ends: every time the catalog grows, or a new category comes in, new cases show up that someone has to resolve. Until now, that “someone” was literally me reviewing pairs one by one, with a model’s help for each hard call. It works, but it doesn’t scale at the pace the catalog is growing.
This milestone was about adding a middle layer: a service built specifically for this kind of fast, high-volume judgment, which resolves the cases where the answer is clear on its own, and sets aside the ones that aren’t. What used to mean looking at thousands of pairs by hand became looking only at the hundreds where there’s genuine doubt — and those I still review myself, with the same care as always. The result isn’t “faster and less reliable”; it’s faster on the obvious, so I can be more careful on what actually needs it.
Still here, still telling the story. What’s next, as I hinted in the last entry, is starting to build the full redesign — this time on top of a catalog that corrects itself faster than I could on my own.