Platform Engineering for Artificial Intelligence (Duy V. Nguyen)（Z-Library）

# Platform Engineering for Artificial Intelligence — Reading Guide ## 【One-Line Pitch】 A practical field manual for platform engineers and ML practitioners who need to move AI from notebook experiments to production-grade systems, covering scalable infrastructure, data pipelines, and full model lifecycle management for generative AI and agentic workloads. Read this if you're tired of broken handshake agreements between data science and operations and want durable patterns instead of silver bullets. ## 【Book Arc】 - **Opening (~0%–9%)**: Establishes why platform engineering matters for AI — the friction of moving from proof-of-concept to production, fragmented toolchains, and the gap between pilot and scale. The preface frames the book's core promise: not a universal blueprint, but durable patterns and trade-offs from real, constraint-heavy environments. - **Early (~9%–33%)**: Covers foundational concepts — defining AI platforms versus traditional ones, data-centric versus code-centric architecture, and the critical planning phase for data pipelines. Chapter 3 is notably deep on pre-design work: OKRs, stakeholder alignment, tool selection, and avoiding ad-hoc pipeline development. - **Middle (~38%–58%)**: Dives into the technical core: architecting data pipelines (batch vs. streaming, Lambda/Kappa architectures, data quality at scale), building modular ML pipelines, embedding governance and security (RBAC/ABAC, lineage, secrets management, bias checks), and implementing Infrastructure as Code with Terraform and Kubernetes. - **Middle (~58%–67%)**: Covers financial management through FinOps (cost visibility, tagging, showback/chargeback) and observability — structured logging, metrics, tracing, anomaly detection, and drift detection for ML models. - **Late (~67%–100%)**: Scales outward: managing platform teams with agile adaptations and intake processes, then applying everything to generative AI — LLM infrastructure, agentic protocols (MCP, Agent-to-Agent), and emerging trends like ethical AI as infrastructure and semantic, graph-native foundations. ## 【Key Takeaways】 - **Platform engineering starts with friction, not technology** (Opening): The moment ML moves from notebook to production reveals missing scaffolding — broken pipelines, model drift, and handshake data contracts. The book's premise is that a coherent platform reduces this accidental complexity. (Early) - **Planning beats ad-hoc pipeline development** (Early): Chapter 3 argues that structured planning — defining OKRs, aligning stakeholders, selecting tools deliberately — prevents costly rework. Key risks include fragmented toolchains, talent silos, and governance gaps that block pilot-to-production transitions. (Early) - **Data pipelines need modularity, fault tolerance, and event-driven design** (Middle): Batch vs. streaming isn't either/or — hybrid Lambda/Kappa architectures balance latency and cost. Data quality must be integrated into pipeline architecture itself, not bolted on afterward. (Middle) - **Governance can coexist with velocity** (Middle): Access control patterns (RBAC as foundation, ABAC for fine-grained policies), end-to-end lineage, automated compliance gates, and bias/privacy testing as code make trust a first-class platform feature rather than a bottleneck. (Middle) - **Infrastructure as Code brings repeatability and auditability** (Middle): Treating compute, storage, networks, and policy as code — via Terraform and Kubernetes — enables consistent AI workloads, least-privilege access, and a pattern library for reusable infrastructure modules. (Middle) - **Cost per model is a measurable, optimizable metric** (Middle): FinOps integrates cost accountability into engineering workflows through tagging discipline, layered tool stacks, and automated optimization levers — making financial management part of the platform lifecycle, not an afterthought. (Middle) - **Observability is a cross-cutting requirement, not a feature** (Middle): Structured logging schemas, fast queryable metrics, trace-log correlation, and drift detection with shift-left strategies are essential for diagnosing performance regressions in AI platforms. (Middle) - **Generative AI shifts platforms from prediction to creation** (Late): A thin global control plane with regional compute planes supports training, fine-tuning, and inference of large models. Agentic protocols (MCP, Agent-to-Agent) and retrieval-augmented generation require new infrastructure thinking. (Late) ## 【Reading Tips】 - **Skim the case studies first**: Each major chapter ends with a real-world scenario (Kyber appears repeatedly — demand forecasting, GDPR compliance, FinOps, IaC transformation). These ground the patterns and show how principles manifest under pressure; read them before the theory if you prefer applied learning. - **Deep-read Chapter 3 (pipeline planning) and Chapter 6 (governance)**: These are the densest with actionable frameworks — OKR management, stakeholder alignment techniques, access control patterns, and compliance gate execution flows. They're the book's intellectual core. - **Treat Chapter 14 (GenAI) as a capstone, not an introduction**: It assumes you've absorbed the earlier platform patterns. If you're primarily interested in LLM infrastructure and agentic workflows, at least skim Chapters 4–8 first for the vocabulary. - **Watch for the recurring Kyber case study**: Following this single company through multiple chapters (streaming pipelines, GDPR, IaC, FinOps, demand forecasting) shows how platform principles compound over time — a useful narrative thread. - **The excerpts don't cover every chapter in depth**: Chapters 9 (hybrid tooling), 11, and 13 (scaling, disaster recovery) appear only as fragments. If those topics are critical to you, verify coverage before relying on this guide. ## 【Coverage Limits】 This guide synthesizes roughly the first 71% of the book in detail; later chapters (hybrid tooling stacks, platform team scaling specifics, disaster recovery rehearsal, and full GenAI/agentic protocol details) are only partially represented in the source excerpts. Specific code samples and tool-by-tool comparisons are not covered here. ##