Shopping by body shape, without ever sending body data to the browser

The problem

The idea is easy enough to explain: you scan your body with your phone’s camera, find out which styles suit your shape, and see real products that match, either in the platform’s own app or on a retailer’s website.

There are two hard problems behind that. First, every product in every retailer’s catalogue has to be tagged against a detailed style taxonomy (fit, neckline, sleeves, length, fabric and more), and catalogues do not arrive that way. Second, body measurements are sensitive personal data, and retailers needed the benefit of that data without ever receiving it.

The risk

A quick build would ask a language model to tag each product and take its word for it. Models are inconsistent, though: tag the same product twice and you can get two different answers, and now and then they will state things that are simply not on the page. At catalogue scale nobody would notice, right up until shoppers were being shown clothes that do not suit them.

The obvious retailer integration is also the dangerous one. If you send the shopper’s body profile to a script on the retailer’s page and filter there, you have put personal data into someone else’s browser environment, where nobody is in control of it.

The decision

We kept AI where it earns its place and left it out where it does not. Body shape classification is a deterministic rule engine, which can be tested and explained, with a model used only for the one rule that needs judgement.

Product tagging is constrained to the taxonomy, with guardrails so that the model can only cite facts that are actually on the page. Prompts live in the content system, so they can be changed without a redeploy. We also measure quality rather than hoping for it, through regression tests against baselines, consistency tests that tag the same product repeatedly to see how much the answers vary, and human reviewers who vote on the AI’s tags. Those votes feed straight back into improving the prompts, so the people who know clothes stay in charge of what “right” looks like.

For retailers, all the filtering happens on the server. A shopper confirms who they are by scanning a QR code with the app, and the retailer’s page only ever receives a filtered list of products. The embed is small (roughly 1 KB to load and 11 KB for the interface, compressed) and isolated, so it cannot clash with the host site.

The outcome

There is now a working pipeline from a body scan to a shortlist of suitable products, in the consumer app and on any retailer’s site, and body data never leaves the server. As catalogues grow, the quality of the AI tagging is tracked and improved through expert review.