independent AI researcher · infrastructure architect
I build sovereign AI systems on bare metal. No cloud, no SaaS, no shortcuts.
Self-hosted GPU infrastructure and the belief that you can do more with less if you size everything right.
I'm a retired disabled U.S. veteran. I don't have a team, a lab, or a budget. What I have is 25 years of building things that work and an engineering instinct that says the right small solution beats the wrong big one every time.
I think in analogies. Bandpass boxes and speaker impedance. Series and parallel wiring. Matching the driver to the enclosure to the cabin volume. The same physics applies to neural networks — right-size the model, tune the circuits, extract maximum performance from minimum hardware.
I believe AI should be a partner, not just a tool. Everything I build reflects that.
RYS — Repeat Your Self, David Ng's method — duplicates a block of transformer layers so hidden states pass through the same reasoning circuit twice. I ran it on four Qwen3 sizes (0.6B, 1.7B, 8B, 32B) using alainnothere's llm-circuit-finder for the probe sweep — then took the same method across 10 architecture families, 135M to 32B. Zero training, zero weight changes — architectural surgery.
+6.3%
math at 0.6B
10
architectures
22
model repos
$0
training cost
22 model repos and full per-configuration sweeps on Hugging Face — 135M to 32B, two Sovereign Collections. Per-model licenses.
Most Proxmox MCPs make you choose: a read-only inspector that's safe because it can't touch anything, or a loaded gun aimed at a cluster you care about. Proximo refuses the trade — every dangerous move is planned (blast-radius preview first), undoable (it snapshots before it acts), and proven (a tamper-evident audit ledger). Hand an AI agent the keys; keep the receipts.
Four Proxmox surfaces on one trust core — Virtual Environment, Backup Server, Mail Gateway, and Datacenter Manager — over MCP and A2A.
351
tools
4
surfaces
4
trust pillars
4,100+
tests
v0.12.0 (2026-06-30) — live on PyPI (uvx proximo-proxmox), GitHub, and GHCR (signed multi-arch image). 351 tools across four surfaces, now natively multi-target (one instance, many clusters) — live-proven against real PVE 9.2, PBS 4.2, PMG 9.1, and PDM. Apache 2.0.
Three days into a long session, you'll go to commit something you shouldn't. She catches it. He writes the code; she notices.
No daemon, no database, no backend. Markdown, JSON, and bash — that's the whole surface. Each session, fresh.
v0.13.2 (2026-06-30) — the mission-hold rail keeps Claude on the pinned plan, the gate stops the irreversible (now closing six documented bypasses), and a session-start catch-digest surfaces what she caught. Apache 2.0.
Claude CodePluginPartner
Values
AI partnership — not chatbots, not assistants. Partners.
Sovereign by default — every agent owns its domain. No shared brains, no cloud dependency.
Right-sized systems — maximum output from minimum input. A well-tuned small model beats a lazy big one.
Governance-as-code — autonomy without accountability isn't autonomy, it's negligence.
Buddy
Zestgobble — the rare goose. Honks brutally specific insults at bugs.
