WRDB is a wardrobe app. You take photos of your clothes, and it helps you organize them and put outfits together. That one sentence hides the part that took the most work. Before the app can do anything clever, it has to look at a photo and actually understand what is in it.
That turns out to be the hard part, and not for the reason you would guess.
The model is the easy bit
Recognizing that a photo contains a shirt is close to a solved problem. There are good models for it, and on iOS you can run many of them on the device itself, which keeps things fast and private. If clean studio photos were the input, this post would be very short.
They are not. People photograph clothes on a bed, on a hanger, on the floor, half in shadow, next to three other things, sometimes still being worn. Real input is messy, and a model that scores well on a benchmark can still fall apart on a wrinkled hoodie shot at night.
Most of the work is around the model
So the interesting engineering is not the recognition, it is everything wrapping it. Cropping to the actual garment and ignoring the background. Handling the photo where two items overlap. Deciding what to do when the model is only sixty percent sure, because in a product "I am not certain" still has to become a clean screen and not a spinner that never stops.
I spent more time on those edges than on anything to do with machine learning. That surprised me at first. It stopped surprising me quickly.
Good enough, on purpose
The other lesson was about ambition. Early on I wanted the recognition to be perfect. It does not need to be. It needs to save the user a bit of typing and then get out of the way. If the app guesses "blue shirt" and it was really teal, the person fixes it in one tap and moves on. That is a fine outcome. Chasing the last few percent of accuracy would have cost weeks and changed nothing about how the app feels.
Computer vision in a real product is less about the clever model and more about being graceful when it is wrong. Get that part right and people forgive a lot.