Case study
Receipts Discovery
Choosing a beauty provider means judging whether their work fits what you want. Receipts lets people search with words or images and see what sits behind a recommendation.
Problem
A polished profile, a similar-looking image, and a review from a client tell you different things. Search needs to use those signals without presenting them as interchangeable evidence of what a provider can do.
My contribution
| Area | What I built and why |
|---|---|
| Name the source | Built distinctions between provider claims, reviews, media evidence, and moderation. A result can identify the kind of support behind a match. |
| Understand the request | Built search around words, images, location, and saved looks. People can express what they want without needing the exact label used by a provider. |
| Explain the match | Built ranking and result explanations that carry the source and reason for a match. Visual similarity can be labelled separately from support in a portfolio or review. |
Decisions
| Choice | Why it matters |
|---|---|
| Keep a claim attached to its source | The result model carries an evidence source and an admission reason. Those distinctions survive into the public explanation instead of being collapsed into a single confidence score. |
| Treat an image as a search clue | A reference image can express a desired look. Similarity helps find candidates, but it does not by itself establish that a provider has done that work. |
| Return the explanation with the result | Search returns labels for different match paths, including visual similarity and portfolio or review support. The interface does not need to invent an explanation after ranking. |
- Provider claimWhat the provider says they offer.
- Portfolio or reviewWork or an account of an experience that supports the match.
- Visual similarityA reason to look closer, with its own label.
What it enables
Receipts explains why a provider appears in search and identifies the source behind the match. People can distinguish a visual resemblance from support in a portfolio or client review when comparing providers.
Lesson
A ranking score is useful to a retrieval system, but people need an explanation they can judge. Keeping the source visible forces the product to be precise about what it knows and what it is only suggesting.