How we taught computer vision to find lost pets
In the Find a Friend project, we chose not to use standard classification, but similarity search instead: we turn a photo into a vector, compare it with cosine similarity, and look for that exact pet among real-life changes in appearance.
In the Find a Friend project, we went with similarity search instead of standard classification. Why? Because real life is nothing like ideal Kaggle images: a pet can lose weight, get dirty, change angle, and generally look after an escape like it just came out of a hackathon 😅
What we built: photo → vector → cosine similarity / ANN.
We trained the model on a domain-specific dataset: breeds, coat color, unique markings. Then we don’t look for “this is definitely a corgi,” but for how similar this pet is to the one we’re looking for.
To keep search fast at real-world scale, we use approximate nearest neighbors. And when there’s a match within the right radius, we send push notifications — while everyone argues about whether AI will replace everything, we’re happy about something more down-to-earth: hundreds of reunited pets among 50,000+ users.
When CV solves not a vanity problem, but actually helps bring a friend home, that’s no longer hype — that’s good tech.