Skip to content

Chapter 9.2 — Minimum bibliography by track

  • Traditional ML / MLOps: MLflow, Feast, Evidently, scikit-learn docs; Mitchell et al. — Model Cards for Model Reporting; Gebru et al. — Datasheets for Datasets; Hardt, Price & Srebro — Equality of Opportunity; Barocas, Hardt & Narayanan — Fairness and Machine Learning; ICLR 2025 blog on calibration.
  • LLMs / Fine-tuning / Quantization: Vaswani et al. — Attention Is All You Need; Sennrich, Haddow & Birch — BPE; Hu et al. — LoRA; Dettmers et al. — QLoRA; Frantar et al. — GPTQ; Lin et al. — AWQ; Xiao et al. — SmoothQuant; bitsandbytes and llama.cpp docs.
  • Embeddings and retrieval: Kusupati et al. — Matryoshka Representation Learning; OpenAI/Cohere/Nomic embeddings docs; Malkov & Yashunin — HNSW; Faiss docs; Pinecone/Weaviate/Qdrant/OpenSearch docs; Microsoft GraphRAG; Neo4j GraphRAG.
  • Serving / Throughput: vLLM (PagedAttention, continuous batching); TensorRT-LLM (in-flight batching, paged KV cache); Hugging Face TGI; speculative decoding (Leviathan et al.); BentoML — prefill/decode.
  • Agentic AI / MCP / A2A: MCP spec; A2A spec (a2a-protocol.org); Anthropic engineering blogs; OpenAI Agents SDK; LangGraph; AutoGen.
  • AI gateways: LiteLLM, Kong AI Gateway plugins, Portkey, Cloudflare AI Gateway (official docs).
  • Security, identity and authorization: OWASP LLM Top 10; NIST AI RMF; NIST GenAI Profile; MITRE ATLAS; OPA/Rego; Cedar; SpiceDB; OAuth 2.0 (RFC 6749), Token Exchange (RFC 8693), Device Grant (RFC 8628), PKCE (RFC 7636), OAuth Security BCP (RFC 9700); MCP Authorization.
  • Observability: OpenTelemetry GenAI semantic conventions and GenAI agent spans.
  • Supply chain: SLSA v1.2, SPDX, CycloneDX (incl. ML-BOM), OpenSSF Scorecard, Sigstore/cosign, in-toto, NTIA/CISA SBOM guidance.

The full bibliography, with URLs and thematic organization, is in references.md.

🏁 Closing

If you made it this far, three ideas — all variations of the same thesis — should have become firmer:

  1. AI systems in production are not just models. They are sociotechnical systems that must be designed, operated, observed, audited and deprecated as critical software.
  2. Deterministic control is not probabilistic mitigation. Where the risk is real, the defense must be in code, authorization, schema and policy — not in free text inside a prompt.
  3. Most of the real value is in the system, not in the model. Swapping the model rarely solves the right problem; working on the surrounding pipeline almost always solves more than it seems.

The focus was on the common case, not the spectacular: a small team, a real budget, an audit still on its way, a customer who will not accept a made-up answer. If the material helps someone design with more patience, operate with more tools, and stop before declaring "100% safe", it has already done what it set out to do.

📌 Status of this draft

This version is an open technical draft, under construction. In particular:

  • Some parts are more mature than others; uneven depth is expected at this stage.
  • The practical examples referenced throughout the book have not yet been implemented. They are planned in the labs index and will be delivered in later cycles. The text is structured so that this future delivery fits naturally.
  • The "Hermes Logística" case study appears only as narrative, with no associated code.
  • The bibliography is deliberately lean; a round of external review is planned for later.

None of this blocks reading or using the material as an evolving public handbook. What is here is honest about what it delivers.

🛤️ Where this project is going

The planned public path, in order:

  1. Public v0.1. Marks the point at which the material can be shared openly as a technical handbook under construction, with no promise of completeness. It is the version this stabilization round is pointing toward.
  2. Implementation of the practical examples (labs/), following the matrix and order in the examples map.
  3. Publishing pipeline (GitHub Actions -> PDF on each release).
  4. Derived posts guided by the internal posts map.
  5. External technical review via community issues and PRs.
  6. v1.0. A coherent set of edited text + runnable examples + completed external technical review.

Each step is incremental. Nothing here depends on a "perfect launch": the material matures in public cycles.

🤝 How to follow and contribute

  • Issues and PRs are welcome: a technical error, an important omission, a better way to explain a concept, a missing primary reference. The rules are in CONTRIBUTING.md.
  • Discussions about editorial direction are better in an issue than in a direct PR.
  • Do not send marketing. The project is deliberately neutral; promotional PRs are refused with an explanation.

To follow the evolution outside the repository:

🪦 About what did NOT make it in

This book deliberately does not cover:

  • Training a foundation model from scratch.
  • Deep detail of RLHF/DPO/KTO beyond what is needed for vocabulary.
  • Accelerator hardware at the microarchitecture level.
  • An exhaustive catalog of the framework of the week.

These topics have specialized material better than this book would produce. The focus here is AI engineering in production, not research nor a framework manual.

🧠 Final reminder

AI systems in production are not just models. They are sociotechnical systems. Treat them as such.

Happy reading, good engineering, and beware of absolute promises.

  • Ruan Pato