Hardcover · 214 pages · $88
Computational Sovereignty
Building Local-First AI Systems
Twenty-four chapters on building AI systems that run on hardware you own. Models, retrieval, knowledge graphs, governed agents — each one measured, each one reproducible, none of them rented. The cloud is treated as what it is: a liability to be priced, not a default to be accepted.
Drag to rotate — it runs on your machine
01 — What you will have built
Six systems, running on your own machine.
Not understood. Built. The book is organized around working systems, in the order you would actually construct them.
A working inference stack
Open weights downloaded, quantized, and served from your own machine. You will know what the GPU is doing at every step, because you configured it.
Retrieval you can defend
Embeddings, vector search, and a RAG pipeline evaluated against a frozen gold set. Recall is a number you measured, not a feeling you have.
A compiled knowledge graph
Documents turned into entities and relations, reasoned over at compile time. The answers were organized before the question was asked.
Agents that obey
Tool use through state machines, watched by an auditor pattern. No vibes, no silent failures, no agent improvising with your filesystem.
Routed persona systems
Quantified personalities and MoE-style routing between specialized local models, running in production without phoning home.
The verification discipline
Refusal logic, contradiction detection at scale, frozen benchmarks, a security model for air-gapped machines — and how to ship all of it.
02 — The instrument
The numbers are the point.
A book that insists on measurement should be willing to measure itself. These are the specifications, not the marketing.
03 — The writing contract
How this book is written.
Four rules the manuscript obeyed for 46,000 words. You can verify all of them against the sample chapter.
Every claim has a number
Recall percentages, parameter counts, VRAM budgets, dollar costs. If a statement can't be measured, it doesn't survive the chapter.
Every pipeline has a repo
Nothing in the book is a screenshot of a demo that worked once. The systems described are built from components you can clone and run.
Every trade-off is priced
Local-first AI costs hardware, attention, and a capability ceiling. The book lists the price before it sells you the machine.
No cloud required
All 24 chapters assume a machine you own and a network you can unplug. The cloud appears only where it is honestly the right call — priced as a liability.
04 — Honest filter
Who this will annoy.
A book with a thesis should say who the thesis is against. This one does.
Cloud-first teams
If your architecture diagram is a list of vendor logos, chapters 1 and 22 will read like an intervention. That is intentional.
Demo builders
If your standard of done is a video that plays well on social media, this book's insistence on frozen benchmarks will bore you. Good.
Prompt collectors
There are no prompt packs here. There are state machines, evaluators, and audit ledgers. The magic words are load-bearing code.
05 — Research log
Tested in public first.
The open notebook behind the book. Every technique in these chapters was built, measured, and argued about here before it earned a page.
Hardcover · 214 pages
Read the first chapter. Then decide if you want the other twenty-three.
Chapter 1 is free, ungated, and representative. If its measurement discipline doesn't convince you, the book won't either.