AI Sizing Assistant

A Medium in one brand is a Large in another, and returns driven by wrong sizes are a persistent cost for online fashion. Most sizing tools attack this by asking shoppers to measure themselves, a step most people skip or get wrong.

The product needed to turn two to four 'this fits me well' data points into a usable fit profile, accept a size question in whatever form the shopper has it, a typed brand name, a pasted product link, or a photo of a size chart, and give an honest answer: when no size in a brand truly fits, say so and offer the nearest option rather than fake confidence.

It also needed to give the operating team full control of the sizing dataset, since recommendations are only as good as the brand data behind them.

The size recommendation itself is computed by a deterministic engine: anchor ranges are intersected into a centimeter fit band, matched against each brand's size rows, then adjusted for fit preference and known brand bias. AI models are used only where language and vision are genuinely needed, interpreting a free-text query, reading a product page, parsing a size-chart photo, answering follow-ups, and are explicitly instructed never to alter the engine's recommendation. Every AI path has a non-AI fallback.

Honesty is product behavior, enforced server-side. Ambiguous brand queries return candidate lists to confirm rather than a guess; blurry size-chart photos are rejected with a request for a better shot; a chart that visibly belongs to a different brand is flagged; and when the profile lands outside every size range, the result is marked no-fit with the nearest size and whether it will run snug or loose.

We consolidated the originally specified separate backend into Next.js route handlers backed by Supabase, one deployable unit per app, cookie sessions, and Postgres row-level security policies doing the enforcement work a separate API layer would otherwise carry.

Onboarding is a guided five-step flow that resumes exactly where the user left off. The anchors step is the heart of it: the user searches the brand list, picks the category and the size that fits them well, with two anchors required and up to four sharpening the profile. Body measurements are an optional refinement, never a requirement, and the user never sees raw numbers.

To check a size, the user types a brand and item, pastes a product link, or uploads a photo of a size chart. The result shows the recommended size in large display type, an alternative, a low/medium/high confidence indicator, and a one-line reason referencing one of the user's own anchors. An AI chat box answers follow-up questions about that specific result, and every past check is kept in a reopenable history.

The commercial wrapper is complete: a metered free plan with capped checks and AI messages per rolling 30-day window, Stripe hosted checkout and Customer Portal, and webhook-driven subscription lifecycle handling including renewal-time price migration. The admin panel gives the operating team brand records with bias flags and verification status, per-gender size rows in centimeters, an Excel bulk importer with per-row reporting, pricing configuration that drives both the marketing page and Stripe, review moderation, and admin invites.