Automated Product Matching Pipeline

Digital shelf analytics sounds simple until you try to compare "the same product" across two retailers and realize there's no shared ID connecting them. One site calls it "Men's Classic Leather Sneaker - White," another calls it "White Leather Trainers, Men"; the photos are shot differently, and the size and color attributes are structured differently too. A price comparison built on that mismatch isn't comparing the same product at all.

The matching layer is the unglamorous, load-bearing piece underneath every other kind of retail intelligence: pricing, assortment gaps, competitor benchmarking are all only as trustworthy as the matching beneath them. So I built a pipeline that reconciles listings across competing fashion retailers despite inconsistent titles, images, and attributes, turning approximate comparisons into genuinely apples-to-apples ones.

A false match silently corrupts everything built on top of it. A missed match is just a gap.

That asymmetry shaped nearly every decision in the matching logic. It's tempting to tune a matcher for the highest overall accuracy, but the cost of a false positive and a false negative aren't the same, so the system deliberately errs toward caution: when it isn't confident two listings are the same product, it says so, rather than guessing.

The tradeoff is a familiar one to anyone who's built entity resolution: optimizing for accuracy on paper isn't the same as optimizing for the outcomes that actually matter downstream.