The real problem: recognizing the same product
Four supermarkets, each with its own way of naming and packaging the same item. The story of Nada Menos's hardest problem, and a catalog that broke once.
This is the heart of the project, and the milestone with the longest story behind it. Four different supermarkets, each with its own way of naming, photographing, and packaging the same products. A six-pack of one-liter water bottles and a case of twelve half-liter bottles aren’t “the same product” in total volume, even though the price per liter can fairly be compared between the two. A “pack of 3” is sometimes three separate units and sometimes a single larger unit that bundles three. And someone searching for a specific milk brand expects to see the real presentations that exist — six, seven of them — not a separate row for every way each store happened to write the name in its own catalog.
It took weeks of working through this, with more than one attempt that had to be scrapped entirely once the results didn’t hold up at scale. At one point during this stretch, a rule that looked reasonable on small examples stopped being reasonable across the thousands of real products in the catalog — and the whole catalog broke for a while. It was the first time the project ran head-on into something this whole way of building has to learn sooner or later: looking fine on three examples proves nothing about the ten thousand cases you didn’t look at.
Out of that came a discipline that stuck for good: measure with hard numbers before calling a change good, never trust “looks fine” on its own.
With the catalog reliably recognizing its own products, the next problem was different: what good is knowing something is cheaper at another store if the extra trip to get it costs more than what you’d save.