AI Engineering in Practice (MEAP) (Richard Davies, Rafael Fischer)（Z-Library）

【One-Line Pitch】 A practical field guide to turning ad-hoc prompting into dependable production systems: it teaches prompt design, RAG grounding, routing, validation, and agent architectures for anyone building real applications on LLMs. Best for developers, product engineers, and technical leads who need reliability, cost control, and measurable quality—not just clever one-off prompts. 【Book Arc】 - **Opening (~0%–10%)**: Frames the core problem—ad-hoc prompting collapses at production scale—using contrasting cases (a hallucinating airline chatbot vs. a high-volume engineered assistant) and introduces the five-layer blueprint: routing, RAG, prompting, validation, and integration. - **Early (~10%–28%)**: Walks through a single customer transaction to show how the layers interact, then pivots to foundation-model fundamentals: why they are called foundation models, their applications, and the architectural logic behind reliable systems. - **Middle (~28%–47%)**: Explains how models are built and behave—data collection and preprocessing, Transformer self-attention, training objectives, instruction fine-tuning, evaluation methods (perplexity, human ratings, benchmarks), and inference-time generation controls. - **Late (~47%–60%)**: Covers distribution and operational trade-offs: commercial APIs vs. open-source deployment, quantization, inference engines, LoRA fine-tuning, token costs, latency, observability, and security risks like prompt injection. - **Ending (~60%+ of indexed excerpts)**: The excerpts do not cover the later chapters in detail; the book's stated structure promises advanced material on chaining prompts into workflows, intelligent routing, RAG implementation, autonomous agents, and operational practices. 【Key Takeaways】 - **Production reliability comes from layered defenses, not better prompts alone** (Opening): routing, RAG grounding, validation, and integration convert isolated model calls into dependable systems. - **The demo-to-production gap is an engineering problem** (Opening): single-case success fails at scale, exposing edge cases, context limits, cost explosions, and security vulnerabilities invisible in testing. - **Prompt routing is a major cost lever** (Early): classifying query complexity and sending simple requests to lightweight models can cut monthly API bills dramatically compared to sending everything to premium models. - **RAG grounds responses in verified knowledge** (Early): vector databases retrieve semantically relevant documents—not just keyword matches—and inject them as authoritative context, sharply reducing hallucinations. - **Foundation models are general-purpose bases, not task-specific tools** (Early): a single pretrained model can be specialized, extended, or prompted into many different applications without starting from scratch. - **Model behavior is shaped by training and inference choices** (Middle): data quality, architecture, context length, parameter count, token budget, and generation parameters all affect what you get back. - **Evaluation needs multiple methods** (Middle): perplexity catches regressions, human ratings guide quality, and benchmarks like HELM and MMLU compare models across domains. - **Deployment involves real trade-offs** (Late): APIs offer ease and state-of-the-art access; open-source models offer control at the cost of complexity, with quantization, inference engines, and observability tools as key operational concerns. 【Reading Tips】 - **Deep-read the opening chapter**: the five-layer blueprint and the transaction walkthrough are the mental model for everything that follows—skim later if you already know foundation-model basics. - **Skim the model-fundamentals chapter if you have ML background**: the training, preprocessing, and Transformer material is useful context but not the book's core value. - **Treat the architecture layers as a checklist**: as you read, map each technique (routing, RAG, validation, integration) to your own system's weak points. - **Pay attention to cost and latency examples**: the routing math and operational trade-offs are where engineering discipline translates into real savings. - **Note the Python code examples**: the book assumes no AI/ML expertise, so hands-on readers can follow along without special setup. 【Coverage Limits】 This guide is based on stratified excerpts covering roughly the first half of the book; the later chapters on RAG implementation, autonomous agents, and advanced operational practices are referenced but not detailed in the available material.