The Engineer
Who Explains
Eugene Yan arrived at machine learning the way some people arrive at a second language - by necessity, with enough stubbornness to become fluent. His undergraduate degree at Singapore Management University was in psychology and organizational behavior. His senior thesis was titled "Competition Improves Performance." The irony of that title has aged extremely well.
The career started at IBM, where he worked on workforce analytics and fraud detection. Then Lazada, the Southeast Asian e-commerce platform that Alibaba would acquire mid-tenure. He came in as a data scientist and left as VP of Machine Learning, having built what his team described as the strongest data science organization in the region. Product ranking improvements of 5-20%, push notification click-through rates up 10%, classification models that cut costs by 90%. The numbers moved because the systems worked.
The pivot to healthtech was characteristic of his approach: a Series A startup, Southeast Asia's largest healthcare provider, and four cost prediction models delivered in three months. When the project required Python, and Eugene had never written production Python before, he taught himself enough in a few weeks to ship. That is the Eugene Yan pattern - identify what needs to happen, close the gap between what you know and what you need to know, then do the work.
Amazon came next, and stayed for five years. As an Applied Scientist in the Search and Books organization - the Kindle team, specifically - he built real-time retrieval infrastructure, bandit rankers, recommendation systems, and eventually LLM-powered features for summarization, translation, and Q&A. The engineering problems at Amazon scale are different not just in magnitude but in kind. You are not optimizing for clever solutions. You are optimizing for systems that hold.
By 2023 he had been promoted to Principal Applied Scientist. The principal track at a large tech company is the technical equivalent of staying a practitioner while everyone else is being asked to become managers. Eugene has written about this track at length - the different flavors of principal work, why no single flavor is superior, and why the most important thing is finding the version that matches what you are actually good at. Connect the dots.
In 2025, he joined Anthropic as a Member of Technical Staff. His focus remains what it has always been: building AI systems that work reliably at scale, and bridging the gap between what academic research proves possible and what production systems require. Anthropic's mission around AI safety aligns with a concern that has run through Eugene's work for years - not as an abstract philosophical question, but as an engineering constraint.
The writing is inseparable from the work. Eugene has published over 209 posts on eugeneyan.com, totaling more than 420,000 words. His newsletter has run for six years and has over 11,800 subscribers. He writes, as he has said, about topics he wants to learn about - which means the blog functions as a public research notebook, a learning system, and an act of generosity all at once. The posts that surface on Hacker News do so because they are precise in a way that most technical writing is not.
The open-source work is equally practical. The GitHub repository applied-ml has become a canonical reference for ML teams shipping models in production - a curated collection of papers and engineering blog posts from companies describing real systems. applied-llms.org, co-created with collaborators in 2024, documented lessons from a year of actually building with large language models, before the hype cycle had fully resolved into practice.