WorkSkillsRecognitionAboutContactRésumé ↓
Case Study 03 ยท AI Product Prototype

FitSense

An AI fit-advisor that answers 'will this actually fit me?' at the moment of purchase, not after a return.

Status
Live
Type
AI product, prototype
Built on
Lovable

The problem worth solving

Fashion e-commerce has a structurally broken moment: the buyer is one click from purchase and has no reliable way to answer the only question that matters, will this fit. Size charts are static, reviews are noise, and the real feedback loop is the return. Returns are the single biggest margin leak in the category, and almost all of them trace back to fit uncertainty that nobody resolved before checkout.

The decision: solve it at the PDP, not after

The design bet is about where the intervention lives. FitSense puts the fit decision on the product detail page itself, the moment of highest intent, rather than downstream in returns or upstream in a separate sizing quiz the buyer will not take.

A fit answer is only valuable if it arrives before the decision, not after the package.

The tradeoff I took a side on

The hard tension in any fit advisor is confidence versus honesty. A tool that always projects certainty drives more add-to-carts short term and more returns long term. A tool that hedges on everything is useless. The position I took: the advisor should be confident where the signal supports it, and explicit where it does not, because a fit tool that quietly overstates certainty destroys the exact trust it is meant to build.

What it is

A working, interactive prototype: an AI fit-advisor built into a realistic product-page shell where a user engages with the recommendation directly. It reads the garment size chart, reasons over a fit profile with a live LLM, and returns three things: a size, a one-line reason, and a confidence level. At cold-start, with no history, it says so plainly rather than inventing certainty.

Naming exactly where the AI can and cannot reach is the product judgment this project is built on.

Responsible by design

The spine of the plan is the confidence gate. Because the feature uses personal data, body measurements and purchase history, and reasons in natural language, seven realistic failure cases were mapped and owned directly, from confident-wrong recommendations to construction mis-extraction, cross-brand normalization errors, cold-start over-confidence, and graceful degradation at peak load. The point is not to claim the model is safe, it is to name where it is not and design for it.

§

Full case study & deliverables

The work is documented as a six-part product package: a product audit and friction analysis, a full PRD and feature spec, a responsible-AI and risk plan, a prototype and wireframes build note, a metrics and GTM plan, and a technical appendix defending the design under hard questions.

Next case study
Freddy AI, A Strategic Teardown