Patterns of Application Development Using AI (Obie Fernandez)（Z-Library）

【One-Line Pitch】 A practical pattern catalog for developers who want to embed large language models into real applications without fine-tuning everything in sight. Best for working engineers (especially Ruby/Rails shops) who need architecture, prompting, and orchestration patterns rather than ML theory. 【Book Arc】 - **Opening (~0%–10%)**: Sets the mindset shift — LLMs as reusable building blocks, not magic. Covers tokens, modalities, multimodal models, and why you should design for swappable models via gateways like OpenRouter. - **Early (~10%–32%)**: Fundamentals of steering models: base vs. instruct-tuned models, temperature, domain adaptation, prompt engineering with zero/few-shot examples, and the case for RAG over fine-tuning. - **Early–Middle (~29%–39%)**: RAG in depth — knowledge-source preparation, normalization, entity recognition, synonym expansion, chunking (including proposition chunking), and retrieval orchestration with real prompt examples. - **Middle (~39%–52%)**: The "multitude of workers" turn: AI workers as discrete agents (account management, fraud detection, sentiment analysis), function/tool use, and Ruby-specific concurrency realities like the GIL. - **Late (beyond ~52%)**: Excerpts do not cover the later chapters in detail; the sample thins out after tool use and worker orchestration. 【Key Takeaways】 - **LLMs are composable building blocks, not standalone products** (Opening): The book's core thesis is that you can delegate functions you'd never build by hand — the GPT-4-as-Dungeon-Master anecdote is the framing device, not the point. - **Design for model swappability from day one** (Early): Choosing an LLM is rarely permanent; modular architecture lets you upgrade models without redesigning prompts or datasets. - **Prompt engineering plus RAG beats fine-tuning for most apps** (Early): Fine-tuning carries real cost and maintenance overhead; clear context and narrowed paths handle the majority of use cases. - **Retrieval quality is a data-prep problem** (Early–Middle): Chunking along semantic boundaries, overlapping chunks, normalization, and entity tagging determine whether RAG actually works. - **Proposition chunking sharpens retrieval** (Middle): Breaking chunks into self-contained factual statements improves precision over raw semantic chunks. - **AI workers map cleanly onto existing patterns** (Middle): AccountManagerAI, fraud detection, and sentiment analysis workers show LLM calls slotting into the ROP (Railway-Oriented Programming) pattern with `finished` and `notify_human_administrator` escape hatches. - **Tool use extends what a model can do** (Middle): Function definitions like `save_analysis` and `invoke_google_api` let models act on external systems without hand-writing every integration. - **Ruby's GIL shapes your async strategy** (Middle): Thread-based concurrency works for I/O-bound LLM calls but not CPU-bound work — a concrete constraint the book addresses head-on. 【Reading Tips】 - **Deep-read the RAG and worker chapters** (roughly 29%–52%): These are the most concrete, example-heavy sections and the ones you'll return to during implementation. - **Skim the model-selection and tokenization material** if you already know the LLM landscape; it's orientation, not technique. - **Treat the Ruby/Rails code as illustrative, not mandatory**: The patterns (workers, function calls, ROP integration) transfer to other stacks even if the syntax doesn't. - **Watch for the Olympia case study** as the running example — it ties account management, fraud detection, and sentiment analysis into one coherent application. - **Don't skip the concurrency discussion** if you're building anything with multiple AI calls; the GIL caveat will save you debugging time. 【Coverage Limits】 This guide is based on stratified excerpts covering roughly the first half of the book; later chapters on deployment, evaluation, or advanced orchestration are not represented. Specific chapter titles and any content beyond tool use and worker patterns are not covered by the excerpts.