← all posts site ↗

a wardrobe app needs receipts

i tried turning a few years of personal photos into a wardrobe app. the rough pipeline made sense: throw away near duplicates, group similar outfits, ask a vision model what is in the remaining pictures, then recommend combinations from the result. it looked impressive immediately because there was a database, filters, and three clean little outfit cards. that part was easy.

then the app started recommending clothes that do not exist. a vague gray shirt became a specific button-down. cropped shorts became chinos. somebody else in the background could leak into the catalog as if their clothes belonged to the person the app was built for. the model was not malicious or even unusually bad. it was doing what image models do when the image is partial or ambiguous: filling the blank with something plausible.

the fix was less about making the model smarter and more about making every recommendation falsifiable. every garment card now needs a source photo that opens at full size. labels can be edited in place. bad items can be removed. repeated sightings of the same thing get folded together only after colors and categories are normalized, not because the model happened to phrase them similarly.

the useful version of an ai closet is not an oracle that tells you what you own. it is a messy catalog that keeps pointing back to evidence and lets you correct it faster than you could build it by hand. recommendations should feel like a memory jog, not a confident hallucination dressed up as a fit check.