← all projects

SHKAF

2026·an AI wardrobe pipeline·gpt image 2 · pixel attributes · recipe matcher

A production-grade ingestion pipeline for a wardrobe app: a photo of a garment on a bed becomes a catalogue-quality card on white, attributes are measured from the pixels at zero API cost, and an outfit matcher assembles looks from recipes. Its best feature is refusal — instead of forcing a bad outfit it says "no light bottom for a monochrome look".

in: a photo on a bed
in: a photo on a bed
out: catalogue card
out: catalogue card
assembled look
assembled look
4pipeline stages: restore → cut → measure → match
$0per attribute and per look — computed from pixels

Transferable know-how

The prompt system documents verified API behaviour rather than folklore: the generations endpoint silently ignores reference images while the edits endpoint respects them; every concrete noun in a prompt is a seed; a print is detected by chroma-to-lightness ratio because shadow multiplies brightness while paint changes chroma. The same pipeline later shipped catalogue cards for a streetwear brand and product cards for e-commerce clients.