How we use AI
Last updated: November 2026
No model answers you when you use this site. AI works behind it instead, on the catalog and some of the writing, and this page covers each place it runs: what it does there, and whether a person checks the work. In a few places, no one does yet.
Nothing you click calls a model
The finder, search and the bag pages all run on plain rules that we wrote. When you answer the finder's questions, each bag gets points for every answer it matches, and the bags with the most points show first. The short reasons under each result come from templates, and search matches the words you type. There is no chat and no assistant, and your answers never go to a model. Even the alt text on photos is plain: it is the bag's name, or a line built from its name, brand and color.
Where AI works: building the catalog
AI works earlier, before you arrive, when we turn brand product pages into the catalog the finder searches. It works in three layers, and each one does a narrower job than the one before.
Layer 1: text models, on every product
Small text models run on our own computers and read every product page we collect. Each one has its own job:
- One reads the page text, title and tags, and decides whether the product is a bag and, if it is, which type.
- Another picks out the lines of the page that talk about the bag, and sorts unusual color names into a color family.
- A third finds weight, volume and size in the text, and pulls out short phrases that appear word for word on bag pages.
No person checks their output one product at a time. Instead, we re-check the products they turn away in batches, and a person decides which ones come back.
Layer 2: an image model, on product photos
Where the text models read words, the image model looks at photos. It also runs on our own computers, and it sees one product photo at a time with nothing else: no title, no brand, no other photos. From that photo alone, it says what the photo shows: the angle, whether it is a studio shot or someone wearing the bag, and what is inside an open pocket.
Once it has labeled a bag's photos, its labels decide which photos the bag page shows first. No person checks those picks yet. What people do check is the model itself: they label photos by hand and compare its labels with theirs.
Layer 3: a large language model, for narrow jobs
For a few narrow jobs, we use a large language model run by an outside AI company. We send it brand photos and drawings, never anything about you.
- It weeds out photos that show a different product. It compares each photo with the bag's main photo and answers same product, different product, or not a product photo. A person confirms every "different product" before that photo comes down, but a "not a product photo" comes down on its own.
- It reads one brand's dimension drawings and turns them into numbers. No person has checked those numbers yet.
- It helps label the photos we use to check the image model, alongside other models. A label goes into that test set when a person confirms it or when the models all agree, and a person checks a sample of the ones they agree on.
We also built a step where it fills in missing specs, but that step is switched off. Nothing in the catalog today comes from it.
Writing
Outside the catalog, AI helps write the articles. Some pieces start as an AI draft and some start as ours. Either way, AI edits every piece against our written rules, and no piece goes live until a person has read and approved it. These are the rules it edits against, word for word:
- Writing stages: map, rough draft, final, and the check at each step
- Least first: add only what answers a question the reader would ask
- Short words: the shortest word that does the job
- Writing mechanics: sentences, filler, claims and formatting
Illustrations
The drawings work the same way: we set them up, AI does the middle step, and a person finishes them. Each one starts from a reference image we pick and a written spec of what it should show. An AI image tool traces the reference into a flat drawing in our colors, and every drawing is checked and fixed by hand before it goes live.
AI assistants that read this site
Some readers reach us through an AI assistant instead of a browser. For those assistants, we publish a plain guide to the site at /llms.txt, a shopping guide, and a read-only way to search the catalog. No model runs on our side for any of this: we hand over the same data you see.
The code
We also write the site's code with an AI coding tool, and a person reviews every change before it ships.
When this changes
Everything above is what runs today, not what we plan. We pick the AI tool that gets each job done, and those picks will change. When they do, this page changes with them.
Resources
- The writing rules: writing stages, least first, short words, writing mechanics
- For AI assistants: /llms.txt, /llms-full.txt, shopping guide