Local-first AI systems
Systems that run on your hardware, and can prove what they did.
I build local-first AI systems: agent runtimes that separate what a model proposes from what the machine is permitted to do, knowledge compilers that keep every claim traceable to its evidence, and creative tooling that verifies what it actually rendered instead of trusting an exit code.
01 — The argument
Three constraints that recur in everything here.
The systems differ in language, domain and maturity. They keep solving the same three problems, which is why they belong together.
The execution boundary
This shape appears in the agent stack, the video studio and the specification engine. It is a diagram rather than generated artwork because the labels have to be correct — a model cannot be relied on to spell “allowlist”, and a diagram that mislabels the architecture is worse than none.
02 — Selected systems
Four entries, and the problem each was built to answer.
Maturity labels describe the code. Dates are the last commit, read from the repository — not a promise about the roadmap.
03 — Ways in
Start from the idea, the system, or the argument.
04 — Research log
Tested in public first.
The open notebook. Every technique in the systems was built, measured and argued about here before it earned a place in one.
05 — Book
Computational Sovereignty: Building Local-First AI Systems
Run serious AI on hardware you own — full applications, no cloud, no exfiltration. It is the long-form version of the argument the systems make: that owning your execution environment is a precondition for trusting what it produces. 214 pages, 24 chapters, $88.
06 — The engineer
Ten years of search relevance, twenty of painting, and the same question in both.
I spent 2010 to 2020 at Amazon in Austin measuring whether information actually satisfied the query — relevance as a measurable property rather than a vibe. Before that and throughout, I painted: figurative oil work, then two-hour films made by wiring televisions through modulators to refurbished computers, one of which screened at Austin Film Society in 2012.
The AI work is the same question one level up. A search result either deserves to be surfaced or it does not. A generated claim either survives its evidence or it does not. Both are answered by inspection, not by confidence — which is why these projects publish their tests, their measurements and their failures.
Self-directed computer science through MIT OpenCourseWare, 2017 to 2021 — the curriculum, not the credential.
What I work on
- Local inference and owned execution environments
- Deterministic governance around model output
- Knowledge compilation, provenance and claim graphs
- Evaluation that measures instead of asserting
Elsewhere