The working Athenaeum for recommender systems, search, and ranking.

Atlas indexes papers that matter to practitioners working on retrieval, ranking, and personalisation. Results return title, authors, venue, and year, and link straight to the canonical source — we index metadata and point you at the original, we do not rehost it.

The engine sleeps when idle, so the first search of a session can take up to a minute. Every search after that is fast.

A technical book on building recommender systems in production: multi-stage architectures, retrieval, ranking, value functions, embeddings, and the data pipelines underneath them. Chapters and the accompanying code are public on GitHub under the MIT licence. No account, no email address.

1. 01Introduction to Recommender Systems
2. 02Multi-Stage Architectures
3. 03Training Data and Evaluation
4. 04Scalable Data Pipelines and Feature Management
5. 05Retrieval
6. 06Ranking Basics
7. 07Second Stage Ranking and Multi-Task Learning
8. 08Value Functions
9. 09Advertising and Retail Media Networks
10. 10Item Embeddings
11. 11User Embeddings

A specialist job search for recommender systems, search, and ranking engineers. Instead of filter checkboxes, you describe what you are after — “MTS roles in adtech, remote, exclude FAANG” — and it works from that. JobLens is a hobby project still in beta, and runs as its own product with its own sign-in.

The intent is a place for people who work on retrieval, ranking, and personalisation to compare notes on the parts that do not make it into papers — evaluation that survives contact with production, serving costs, and the failure modes nobody writes up. It does not exist yet. There is no waiting list, and nothing here collects your email.

Most of what is written about recommender systems is either a paper abstracted away from production or a vendor post abstracted away from the mechanism. This is an attempt at the middle: a book that explains how these systems are actually built, code that runs, a search engine over the research that informs it, and a job search for the people who do this work. It is built and maintained by one practitioner, in the open, one capability at a time.