Goal
The path to "生成AIアプリケーション評価入門"
Here is the reading path leading up to this book, derived from its dependencies and ordered from the fundamentals.
The path so far (10 books)
Why read this first: Once you can improve delivery performance through measurement, you widen your view to the engineering culture and practices that sustain it at scale and over time. Google frames this as 'programming integrated over time.'
Why read this first: Learning an architecture that controls the direction of dependencies makes you want testing principles that maximize the resulting testability. Unit Testing: Principles, Practices, and Patterns defines what a good test is, converting the loose coupling your architecture enables into testing value.
Sources
Why read this first: Once you have the concrete techniques of readability such as naming and formatting, you widen your view to the professional philosophy that runs through daily decisions—DRY, orthogonality, and a stance toward change. The Pragmatic Programmer binds individual techniques into an attitude of why.
Why read this first: Once you have the pragmatic mindset—sharpen your tools, own your judgment—you take in generative AI as a new tool. The ability to verify AI output rather than trust it blindly is the very craftsmanship of the AI era.
Why read this first: After understanding the engineering practices of large organizations, you view how generative AI changes the flow of development with the same discipline. AI is a tool that pays off only on a foundation of review, testing, and design judgment.
Why read this first: After learning to use generative AI as a development tool, you face the next question: how do you measure the quality of an application that itself embeds generative AI? This book covers evaluation methods along the development lifecycle, from LLM-as-a-Judge to security evaluation and AI agent evaluation.
Sources
Why read this first: Once the small TDD cycle makes writing working tests second nature, you advance to principles of what and at what granularity to test. Good tests are resistant to refactoring, do not get in its way, and reliably catch regressions.
Why read this first: After practicing how to grow design through tests, you raise the sufficiency of the test cases themselves with systematic techniques. You turn both wheels—good design and thorough cases—to reduce gaps in testing.
Why read this first: After grasping the qualities of good tests—protection against regressions, resistance to refactoring, fast feedback—you raise the question of what to test with systematic techniques like boundary analysis, equivalence partitioning, and structural testing. Case selection moves from intuition to engineering.
Why read this first: Once you have test design techniques like boundary analysis and equivalence partitioning, you apply that thinking to a new kind of target: generative AI applications, whose outputs are probabilistic and whose behavior is complex. This book covers evaluation methods specific to generative AI, from building an evaluation perspective model to confusion matrices and RAG-oriented metrics.
Sources