Selfwear

Fashion has run in the same order for a hundred years: make first, sell later, write off the rest. E-commerce didn't change that – it is a shelf with a search bar. The best a recommender can do is push the closest match in stock to the front. If your piece doesn't exist, you get the next best thing – and the return that comes with it.

Selfwear flips the order: "Sell first. Then make. Or never make." Clothing that comes into being with a person – from their taste, their body, their hesitation – and is only produced once it's paid for. For the brand: no inventory risk, no markdowns, no structural returns. For the customer: a piece that adapts to them, not the other way round.

The road there runs at two speeds. The full vision – generating instead of choosing – needs on-demand manufacturing. But the same engine can already do something else today: read an existing catalog as if it were made for one person. That is where the work with brands begins.

One engine that reads taste – two ways of answering it.

The core: Empathy Engine

• Parameter Flow instead of a profile — a living description of taste: color identity (distribution across color families with learning rate and decay), shape identity (structural dimensions with mean and variance – variance reveals what really matters), body identity (geometry from a few smartphone photos)
• Six engagement signals instead of a binary click: impression, quality impression (1.5 s dwell), click, hero selection, dismissal, revisit – pauses, hesitation and rejection all count
• Convergence tracking — after stable cycles the system shifts from exploring what might resonate to showing what it knows
• Local-first & privacy by design — the most intimate map of someone's taste stays on their device

Mode 1 — Selfwear: generate instead of choose

• Co-creation inside the brand's codes: its materials, its taste, its promise – directed at one person
• Parameters from browsing + geometry from the scan = a Generative Brief. On hero selection the brief freezes and becomes a production order
• A flexible production cell builds exactly one piece – paid for before it exists. No SKU, no stock, no batch

Mode 2 — Catalog engine: existing stock, read as if made for one (in testing)

• The same engine on a real brand catalog: a pilot with the range of a European fair-fashion label, 2,493 items
• AI-assisted feature extraction from product images and data – every item becomes a point in color and shape space
• Color and shape measured separately (named-block PCA: 6D shape block + 2D color block) instead of blended into one number that was secretly almost all color. Result: the silhouette share of the signal rose from 7.8% to 37.2%, and the quarter of the catalog that used to score as "equally poor matches" for varied taste is fixed
• A product rule follows: color may cross categories (shoe and shirt in the same terracotta), shape stays within its family – no "chunky shoe → chunky coat"
• New product lines (e.g. sneakers) only need a re-fit of the shape block, not a new architecture
• Still open: a qualitative feel test with real feeds before wider rollout