Machine Learning Upgrade A Data Scientists Guide to MLOps, LLMs, and ML Infrastructure (Kristen Kehrer)（Z-Library）

【One-Line Pitch】 A practical field guide for data scientists and their managers who need to move beyond notebook-bound models and ship real LLM-powered applications, covering the full arc from data-centric thinking to MLOps, observability, and cost-aware deployment. Best for working practitioners and technical leads who already know some ML and now need the engineering discipline around it. 【Book Arc】 - **Opening (~0%–10%)**: Frames the core problem — most models never reach production — and argues that modern ML must be treated as a holistic, end-to-end system rather than an isolated modeling exercise. Introduces LLMs as the force reshaping both the power and the complexity of that system. - **Early (~10%–35%)**: Builds the data-centric foundation. Walks through an end-to-end example (a YouTube search agent) to show how components interlock, then drills into unstructured data, vector/graph/relational database choices, embeddings, and data versioning with tools like Comet. - **Middle (~35%–55%)**: Moves into standing up LLMs. Covers the generalist nature of foundation models, the build-vs-buy decision, privacy and PII considerations, and the real cost math behind hosted versus self-deployed inference. - **Late (~55%–80%)**: Excerpts do not cover this stage in detail; based on the book's stated scope, this is where training, fine-tuning, evaluation, and integration of LLM applications into larger systems are addressed. - **Ending (~80%–100%)**: Excerpts do not cover this stage in detail; the book's stated goals point toward monitoring, retraining on new data, and sustaining LLM applications over time. 【Key Takeaways】 - **ML must be viewed as an end-to-end system, not a model in isolation** (Opening): the book's central reframe — data, retrieval, embeddings, UI, and monitoring are interdependent, and a weak link anywhere degrades the whole. - **The production gap is the real problem to solve** (Opening): the excerpts cite a Gartner survey finding only about 54% of models reach production, which motivates the entire MLOps emphasis. - **Data-centric choices drive LLM quality** (Early): database type (vector vs. graph vs. relational), embedding model selection, and chunking strategy are framed as first-class engineering decisions, not afterthoughts. - **Observability and reproducibility are design principles, not add-ons** (Early): tracing prompts, responses, and metadata lets you diagnose quality degradation; versioning data (via artifacts and lineage) enables recovery, collaboration, and auditability. - **Some systems can improve themselves** (Early): retrieval pipelines that accumulate data over time get better with use — a property worth designing for deliberately. - **Foundation models are generalists** (Middle): the best model for one task is often the best for most others, which simplifies model selection but shifts the hard problems to cost, privacy, and integration. - **Build-vs-buy is a cost and privacy calculation** (Middle): the excerpts walk through token-level cost estimation and PII constraints to decide between hosted APIs and self-hosted inference. - **ML infrastructure is now largely buy-and-plug-in** (Early): best-in-class components can be installed or purchased, shifting data scientists' energy from infrastructure plumbing toward application logic. 【Reading Tips】 - **Deep-read the end-to-end example chapters** (the YouTube search agent): they anchor every later abstraction and are the clearest demonstration of the book's systems thinking. - **Skim the tool-specific walkthroughs** (Zilliz, Comet, LangChain setup) on first pass — return to them when you actually adopt those tools, since versions and UIs date quickly. - **Treat the cost and privacy sections as decision frameworks**, not recipes: the token math and PII questions generalize even when your numbers differ. - **Watch for the observability/reproducibility pairing** — these two principles recur and are the most transferable lessons for anyone running LLM apps in production. - **Managers can read the Opening and Middle sections alone** to get the strategic framing; practitioners should read straight through for the implementation detail. 【Coverage Limits】 The excerpts are heavily weighted toward the opening and early chapters (data-centric foundations, databases, embeddings, versioning) and the middle (LLM selection, cost, privacy). The late and ending stages — fine-tuning, evaluation, integration, monitoring, and retraining — are referenced in the book's stated scope but not substantively covered in the available excerpts, so this guide's treatment of those stages is necessarily thin.