Sovereign Local AI: A Bilingual Book on Local AI

LLM
Local AI
Quarto
Python
R
A book in French and English on running AI on your own hardware: 19 chapters in 5 parts plus 5 appendices, with runnable companions: a fine-tuning pipeline, an evaluation gate, a minimal RAG and an Ollama Modelfile.
Published

August 26, 2026

Overview

Running a capable model on your own machine stopped being a hobbyist exercise, and the field moves faster than any book can follow. This one is written for that constraint: the method chapters cover inference parameters, hyperparameters, retrieval and evaluation, while the parts that age, such as benchmarks and product facts, are marked as an early-2026 snapshot with their sources.

It is one Quarto project rendered as one book in two languages, delivered as a readable site, a self-contained HTML file, a PDF, a DOCX and an EPUB.

What is in it

  • Part I: Foundations: why now, choosing a model
  • Part II: Running models: ecosystem and install, Studio Chat, inference parameters
  • Part III: Training: fine-tuning pipeline, hyperparameters, data recipes
  • Part IV: Developing and deploying: export, coding agents, harnesses, Python, RAG cookbook, voice/vision/images, serving a team
  • Part V: Governing: evaluation, GDPR and hardening, business case, strategy paper
  • Appendices: researcher primer, sources and limits, projects directory, hardware guide, DPIA template

Every chapter also renders as a self-contained reveal deck with numbered, shareable URLs, plus its own PDF and HTML page. Companion code ships with each release: a fine-tuning script, an evaluation gate, a minimal RAG, an Ollama Modelfile. Chapter 3 embeds a WebLLM demo that runs in the browser with no install.

How it is built

Chapters are bilingual source fragments (::: {.en} / ::: {.fr}). A pre-render hook regenerates the merged book, the per-chapter documents and the project file from a single order file, so the table of contents cannot drift out of sync; a filter then selects the language, restores canonical heading ids and strips the bilingual speaker notes.

Slides and the book are deliberately separate documents: Quarto renders a document’s HTML formats concurrently, and they race on the shared _files directory that self-contained slides depend on.

CI renders both languages on every push, publishes the site, attaches PDF/EPUB/DOCX to each v* release, and runs a monthly link-freshness audit that opens an issue when a chapter goes stale.

Read it

Back to top