The Tradeoff Method for ML Systems: Machine learning and GenAI system design interviews - Tapa blanda

Sinopsis

Most candidates can name the model. Few can name the objective.

A machine learning design interview does not ask which model to use. It hands you a business goal and a rate, and watches whether you can turn them into an objective a model could serve, find the constraint that binds, and choose a design whose price you can state. The model is the last thing decided, and the interviewer knows it.

The Tradeoff Method for ML Systems retunes the five moves of The Tradeoff Method for systems with a learned component: name the objective, bound the problem, find what binds, choose under it, close the loop. Each move leaves an artefact on the board, and two are new to this volume: the objective sheet, which turns a business goal into an ML objective with its proxies and guardrails, and the loop sheet, which says how the design will be known to work and what it does to its own data. Between them sits the budget triangle, quality against latency against cost, on which every design in the book is drawn.

What is inside

• The method and the recurring tradeoffs: objective, proxy and guardrail; the budget triangle; data and labels; the training–serving gap; model capacity against the clock; serving and unit economics; evaluation, drift and harm.
• Fourteen designs worked end to end: a short-video recommender and a personalised feed; marketplace search, ad selection and delivery-time estimation; card fraud, content moderation and visual near-duplicate search; an enterprise assistant over company documents, a returns agent with tools, a serving platform and an assistant with memory; and two staff-level briefs, a platform consolidation and a forty-per-cent cost mandate.
• The clock: the 45- and 60-minute plans minute by minute, nine recovery moves, a mock-interview protocol, and what the interviewer writes down.
• Rubrics for senior and staff loops, the applied-scientist and platform variants, sixteen red flags, and a self-marking sheet for every design.
• 80 exercises with full solutions, most of them arithmetic with a decision at the end.
• Twelve unworked briefs for study groups, each with a candidate's half and an interviewer's half.
• Reference numbers for learned systems, from tokens per second to label delays, in one appendix.

Who it is for

Engineers and applied scientists who have shipped a model or two and are preparing for a senior or staff ML or GenAI system design loop, who can name the architectures and still find themselves unable to say why this one, at this price, for this objective. It is not a first course in machine learning and not a catalogue of architectures; it cites the canonical papers at the end of every chapter and assumes you know what a ranker, an embedding and a prompt are.

How it is different

Architecture books teach shapes. This book teaches the decision that picks one shape over another, with the arithmetic done on the page: the fleet from the rate, the cost per thousand, the retraining cadence from the staleness curve, the judged set from the regression you need to see. Every worked design says which constraint binds, how far the numbers would have to move for the decision to change, and what the loop will measure after it ships.

It stands on its own; readers of The Tradeoff Method will recognise the five moves and find them asked of a harder object. Companion material, including printable artefact sheets, the ML cost-and-capacity toolkit, evaluation templates, the rubrics and errata, is free at github.com/Anag1982/tradeoff-method.

Avishek Nag is an Associate Professor at University College Dublin. His research is in optimisation and the analysis of tradeoffs in networks, and the method in this book grew out of teaching it to his students.

"Sinopsis" puede pertenecer a otra edición de este libro.