Books

Fragment 1 — Chapter 1, "Introduction": why AI governance is a governance question, not an engineering oneWhen an organization first faces a decision to deploy an AI system — whether a credit-scoring model, an automated resume-screening tool, or a customer-support chatbot — the most common governance mistake is to treat that decision like an ordinary IT project: define the requirements, select a vendor, implement, and hand the system over to operations. Artificial intelligence technologies do rely on software and computing infrastructure, and in that sense they resemble any other IT initiative. But they differ in how they generate risk, and that difference calls for a distinct governance approach rather than a simple extension of familiar project management.Risk-oriented AI governance starts from a different premise: before discussing rollout timelines, budgets, or functional requirements, an organization has to answer the question "what adverse consequences could this system cause, for whom, with what likelihood, and how severe would they be" — and, alongside it, "how much of that harm are we willing to tolerate for the expected benefit." Fragment 2 — Chapter 3, "Trustworthy AI: Seven Characteristics of Trust": why explainability and interpretability are not synonymsConsider an AI system that automatically sorts incoming support tickets by priority. Explainability answers the "how" question: which features of the incoming text (keywords, tone, customer history) contributed to the computed priority score and with what weight — a technical account of the computation mechanism. Interpretability answers the "why" question for this particular ticket in the context of the system's business purpose — that is, whether a high priority score means "this customer is losing money right now" or "this customer is statistically likely to cancel," and whether that meaning matches how a support agent should act on it. The two characteristics support each other but serve different governance needs: explainability matters more to the engineer debugging the model, interpretability matters more to the agent or manager who must act on the system's output without understanding its internal mechanics. Fragment 3 — Chapter 18, "The Running Case": where ISO and NIST illuminate each other's blind spotsPrecisely because NIST explicitly requires, in its own standalone subcategory, a designated authority to deactivate a system, and no direct equivalent exists in Annex A of ISO, an organization that implements only ISO/IEC 42001 without checking it in parallel against the AI RMF risks missing this requirement altogether — it dissolves between the adjacent, but not identical, controls A.6.2.5 (release criteria) and A.3.2 (roles and responsibilities). This is a compelling, concrete example of the argument the book already made back in Chapter 1: the two frameworks do not compete but mutually illuminate each other's blind spots. Fragment 4 — Chapter 18, the book's closing paragraphAI risk management, as this book has shown it, never concludes with a signed document. It concludes — and immediately begins again — with the next MAP cycle, the next internal audit, the next policy review, the next model version, which has to be brought into operation just as thoroughly documented, traceable, and responsibly managed as the one before it.

A from-scratch AI/ML course that treats you like an engineer, not a tourist — three modules covering language-as-numbers, the math foundations, and neural networks, each concept built in raw Python first, then PyTorch, so you always know what's really happening under the hood.

Build AI that keeps working when the cloud doesn't. Local-First AI Engineering shows you how to run capable, private AI systems on infrastructure you control. From hardware and inference to RAG, agents, security and production ops, you'll learn how to build local AI that is fast, reliable and truly yours.

Build AI systems that do more than demo well. This hands-on guide shows you how to engineer autonomous software on Linux that is secure, observable, reliable and ready for production. From agents and memory to threat modeling, OpenTelemetry and Kubernetes, you’ll build a real system while learning what it takes to run AI you can actually trust.

不堆技术名词，也不炒概念。两把尺子：一把横向的「五阶智能成长弧」（知→懂→用→长→合），一把纵向的「AI Intelligence Scale 十级刻度尺」（从表示、预测、泛化、因果，一路到规划、行动、元认知、持续学习、发现、自我改进）。

Strip away the frameworks and an AI agent is a language model in a loop with a list of tools. This short guide builds one in sixty lines, runs it for real, and gives you the vocabulary to judge any "agentic" proposal that crosses your desk.

Hand over the boring parts of your week to AI — no coding required. Five practical jobs, every prompt written out, and an honest guide to what to check before you trust it.

Platform engineering is changing fast, and AI is raising the stakes. This practical guide shows you how to build internal developer platforms that work in the real world, from cloud-native foundations to AI-powered workloads. Follow hands-on examples and a production-grade reference architecture to create platforms that are secure, scalable and built to evolve.

Move beyond AI demos and build agents that can actually run IT operations. This practical guide takes you from MCP fundamentals to production-ready monitoring, incident response, remediation, Kubernetes operations, security and multi-agent systems, with complete runnable code and a strong focus on safe autonomy.

一套普通人也能上手的 AI 研究方法论：用三天时间，把任何一件事查到心里有底。

What happens when AI agents stop working alone and start working as a team?Artificial Intelligence is entering a new phase in which intelligent systems can do more than respond to individual instructions. Multiple specialized agents can collaborate, divide complex tasks, communicate with one another, use tools, evaluate results, and coordinate their actions toward a shared objective.Multi-Agent AI Systems: The Complete Handbook for Building Intelligent, Scalable, and Autonomous Agent Teams provides a practical roadmap for understanding this emerging paradigm.The book begins with the fundamentals of multi-agent systems and explains why collaboration between specialized agents can be valuable for complex workflows. Readers will learn about hierarchical, peer-to-peer, and hybrid architectures, along with roles such as manager, planner, worker, critic, and supervisor agents.It then moves into modern frameworks and technologies, including CrewAI, AutoGen, LangGraph, MetaGPT, LLMs, vector databases, and agent memory systems. Practical chapters explain how to design agent teams, decompose tasks, establish communication protocols, manage shared memory, coordinate workflows, integrate tools and APIs, and recover from failures.Readers will also explore advanced concepts such as dynamic replanning, parallel execution, swarm intelligence, agent debates, self-organizing systems, human-in-the-loop workflows, and multimodal agents.The book goes beyond experimentation and addresses the challenges of deploying multi-agent systems in real environments. Cloud deployment, Docker, Kubernetes, monitoring, logging, scaling, evaluation, benchmarking, testing, and cost optimization are included.Real-world applications demonstrate how agent teams can support software development, research, customer service, content creation, and business operations.Equally important, the book examines AI safety, privacy, security, transparency, governance, alignment, and responsible AI development.Whether you are a student discovering agentic AI, a developer building your first agent team, a researcher exploring collaborative intelligence, or a professional preparing for the next generation of AI applications, this book provides a foundation for moving from individual AI agents toward coordinated intelligent systems.Understand the architecture. Design the team. Build the agents. Coordinate the intelligence.

You already know how to build software.You can design APIs, debug production incidents, reason about latency, operate databases, and ship features that people rely on. But AI engineering interviews ask a different set of questions:How do you design a RAG system that does not confidently invent answers? When should you use an agent—and when is a simple workflow safer? How do you evaluate an LLM feature when quality can decline without an error message? How do you control token costs, defend against prompt injection, and safely deploy a prompt change?AI Engineer – Interview Book bridges the gap between being a capable software engineer and being ready for an AI engineering interview.Learn how to reason about models, retrieval, agents, evaluations, cost, latency, safety, and production failure modes. Practice answers out loud. Learn the trade-offs behind the right answer—not just the vocabulary.Stop preparing for yesterday’s system-design interview. Start preparing to build AI systems that work in the real world.

Three full passes. One focused half-pass. A repeatable way to find patterns without losing the evidence trail.

La IA no toma decisiones: traduce el criterio del profesional que conoce su organización. Primero es el criterio. Hasta después es la herramienta.

The eight volumes before this one each teach a single domain, and each is deliberatelyself-contained. The exams do not ask about domains — they ask about systems, and a system spansall of them at once. "A retrieval system is slow, expensive and occasionally wrong — which do you fix first?" needsretrieval, optimisation and operations together. Every one of the 150 questions here crosses atleast two domains, because that is the specific thing a single-domain book cannot teach you.