How to Build an AI Product From Scratch
A chapter-by-chapter playbook on turning an AI demo into a dependable product, using a single running example, an academic research assistant, to make model selection, evaluation and architecture decisions concrete.
This manuscript works through the practical decisions involved in shipping an AI product rather than a demo, using one example throughout, Arxie, a research assistant that retrieves and reads full academic papers to produce citation-grounded answers, on the stated view that following one example in depth teaches more than several shallow ones. Chapters in progress cover model selection based on product constraints, context engineering and prompt design, evaluation against explicit metrics and release gates, agent architecture, and operating an AI product after launch.
The project is being written and released in stages, with model selection and context engineering the most complete chapters so far, and the remaining chapters available as outlines describing what they will cover; the author updates it in small daily increments and has split the manuscript into separate chapter pages as it has grown.
It is published as a documentation site.