How we taught the service not to confuse “that one” cat with its ginger brother

In Find a Friend, we introduced embedding-based matching: every photo is turned into a vector, and then we look for the nearest neighbors. In other words, we give the algorithm visual memory — a must-have in an era when AI is everywhere, from search to phone cameras.

But real life always gets in the way: two ginger cats in the same neighborhood, same angle, same vibe. So we added ranking to avoid giving people false hope. And if the same found cat is uploaded by multiple users, visual similarity kicks in for deduplication — no duplicates and no chaos in the feed.

Another important layer is temporal decay in scoring: the older the listing, the lower its relevance. Because a photo from a week ago and “seen 5 minutes ago” are, to put it mildly, very different signals.

In short: less noise, more real matches. And yes, cats are still the hardest users.