Technology

How the parts that matter work.

Four decisions carry the draft: the price, what's read from the piece, the cutout, and the placement.

The price

The extreme tenth on each side is trimmed, then the median taken. Absent sold data it draws on active asks, discounted, always a range, never false precision.

Attribution

Material, colour, closure, condition, and an estimated size are read from the photographs. The house is read from a visible label when there is one, and inferred from its design signatures when there isn't, hardware, silhouette, stitching, offered as a lead to verify. Sizing that needs a label stays with you.

The cutout

The background is lifted and the piece squared and centred, a cleaner reading, and an image ready to list on a white ground.

Placement

Category and item specifics are resolved for you; the listing is composed and, only on your word, published. Poshmark and Depop are written but placed by hand.

Under the hood

The machine learning

Three models do the reading, one cloud and two on your own machine. The price, deliberately, is not one of them.

Vision-language model

Reads the piece from its photographs: the house, condition, flaws, an estimated size, and the finished listing copy. It also locates each sellable piece in a full-outfit photo, so a whole look is drafted piece by piece. Runs in the cloud.

Segmentation, U²-Net

A convolutional network that lifts the piece from its background for the cutout. Runs locally.

CLIP and nearest neighbour

Encodes the photo into a vector and matches it against your own reference library to suggest a house. Runs locally, no keys.

Pricing is statistics

The price is a trimmed median of real comparables, not a model. Honest and debuggable where a predictor on thin resale data would not be.

The index verb

The library remembers

Drop your own photographs into refs/ and npm run index embeds them into a small local library. From then on, every draft is matched against what you have already photographed, nearest neighbour over your own pieces, no keys, offline. The same cream bag, two outings apart:

Indexing: the cream shoulder bag saved into the reference library
First outing · indexed
The bag enters the library as reference 01.
Matched: the same cream shoulder bag recognised against the reference library at 0.91
Second outing · recognised
A different dress, a different day; the bag matches reference 01 at 0.91.

Illustrative example of the index verb and the visual match.

In candour

What it won't do

It won't authenticate

It can name the house from a mark or its design, but that is identification, not authentication. For high-value pieces, verify with the house or a service before you list.

Dimensions are estimates

It estimates a size from the piece and its type, useful, but a photograph has no true scale. For an exact fit, measure with a tape before listing.

It won't host images

The cutout is produced; eBay needs a public URL, so you host it.

It won't post to Poshmark or Depop

The copy is written; placing it there is yours.