Principal Data Scientist, Core Shopping Data Science

Every day, hundreds of millions of customers arrive on Amazon's Homepage, search for a product, and land on a detail page. Along the way, some of what we show them is irrelevant — a recommendation that misreads what they want, a widget that repeats what they just saw, a headline that doesn't say what it means, a product that shouldn't have been surfaced at all. We have built the ability to detect these defects at scale using Large Language Models (LLMs), and we report them to Amazon's most senior leadership as a company-level measure of shopping quality. What we have not yet built is the confidence to act on every one of them.

That is the problem you will own.

Today, the most consequential categories of shopping defects — product quality, duplication, staleness, irrelevance — go largely unenforced. Not because we cannot detect them, but because we have not yet clearly defined when they lead to customer dissatisfaction. These are genuinely hard questions. When is a recommendation irrelevant rather than merely unexpected? When are two products duplicates rather than legitimate alternatives? Every answer carries consequences for customers and advertisers.

As Principal Data Scientist, you will own the analytical rigor behind Amazon's store quality metrics end-to-end: how they are defined, how faithfully our LLM implementations execute those definitions, and how accurate the resulting judgments actually are across relevance, presentation, product quality, and duplication. You will go deep enough to inspect individual model judgments and read the edge cases where they break, then come back up to argue definitional questions with the leaders who own the outcomes on both sides. You will also define experiments that will turn judgement into facts backed by data.

…