Image search is now live!
Eugene Yan
A dev built a reverse image-search API that finds products just by photo, no keywords needed. Cool idea, though the live demo's since been shut down to save cloud costs.
Based on reporting by Eugene Yan — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Eugene Yan spent a chunk of 2017 building a product classification API from scratch, and instead of stopping there, he pushed further: could he make a reverse image search engine that finds visually similar products without anyone typing a single word? The answer, for a while at least, was yes — you could upload a photo of, say, your kid's mystery plushie and get back a list of lookalikes pulled from a roughly 200,000-item catalog. Yan himself used it to identify an unnamed stuffed toy as a Domo, something he'd apparently never bothered to Google in years of owning it.
The mechanics are simpler than they sound. A pretrained neural network, VGG16 in this case, converts each catalog image into a feature vector — 25,088 32-bit floats per photo, which is a lot of math to lug around. New uploads get the same treatment, then get compared against the catalog using similarity measures Yan mostly hand-rolled in numpy because he wasn't satisfied with off-the-shelf speed. He's candid that he still didn't get it fast enough, and that storage ballooned to four to seven times the size of the original images once you factor in all those feature vectors.
Where this actually shines is style-based shopping: sofas with a particular dimpled tufting, a jacket cut, a shoe silhouette — anything where "it just needs to look like this" beats typing keywords into a search bar. Yan admits the flip side too. Search by spec, like battery life or memory size, and a reverse image tool is useless; a normal text search still wins there. And with only 200k products behind it, the system sometimes came up empty, or returned matches that were, in his words, merely the least bad option available.
He also swapped Flask's development server for uWSGI plus nginx, letting the site handle more than one visitor at once — a small infrastructure upgrade, admittedly moot given traffic was modest. The project quietly wrapped up not long after: Yan started a computer vision course at Georgia Tech and warned readers he'd have less time for updates. True to that, the image search demo has since been pulled offline entirely, killed to cut cloud costs.
My take — AI-written commentary, not fact-checked reporting
This is a nice reminder that solo builders were doing real, useful applied CV work with off-the-shelf pretrained nets years before anyone was calling it AI at parties. The most honest part of the post is Yan admitting the similarity search was never fast enough and the catalog too small to be reliable — that kind of candor about limitations is rarer now than it should be, when every launch post pretends the demo is production-ready.
Read more about this at: Eugene Yan