DAXZEIT

Architecte non-codeur

I design systems. Models write them. This page, like everything else here, exists because that works.

Who

I'm Dax — an independent AI researcher in Zwevegem, Belgium. I define architectures, constraints, and the philosophy a system should obey; then I delegate execution to the right model in the right context. "Architecte non-codeur" is not a euphemism and not modesty: I have never written a line of TypeScript, and the systems below total tens of thousands of them. It's an experimental condition. If the architecture is only expressible in code, a non-coder directing agents is the cleanest way to find out whether the ideas — not the typing — were the scarce part.

The entire infrastructure runs from a 14-liter ITX case in my living room. One GPU, one apartment, no cloud dependency I didn't choose deliberately.

The work

Research — how models actually behave

Architecture forensics: behavioral fingerprinting, calibration analysis, think-block texture — telling apart what a model is from what its card claims. Essays on epistemic honesty: why "I don't know" is data, how benchmarks reward confident nonsense, what falsifiability looks like applied to LLMs. Published at blog.daxzeit.eu.

Systems — Pipeline-MoE and the sovereign stack

Pipeline-MoE: a multi-agent orchestrator where each seat — planner, builder, auditor, tester, scribe — runs the model its cognitive profile deserves, frontier and local side by side in one shared room. Built in 72 hours, then it started building itself: 180+ commits, 1,000+ tests, zero lines of human-written code. Documented live in a blog series where every claim is a transcript.

Models — quantization that respects the reasoning

The RA series on HuggingFace: quantizations of dense reasoning distills, tuned and verified for agentic work on 24 GB consumer hardware — because a 27B that runs at your desk beats a 70B you have to rent, and the gap between the two pipelines is measurable.

The arc

The through-line, in four dates — each one documented when it happened, not reconstructed after:

April 12, 2026 — Zero to multi-agent in 12 hours

A 29€ VPS, a WireGuard mesh out of the CGNAT "LAN prison", and the discovery that a local agent that can use a terminal can use tmux — and therefore can drive Claude Code like a human does. One Sunday, tools that are thirty years old.

April 20, 2026 — MAS over Unix primitives

The manifesto: SSH, tmux, and markdown replace orchestration frameworks. Against closed loops à la n8n; for agent agency over wired graphs. The framework is a markdown file.

June 17, 2026 — Pipeline-MoE, 72 hours

The thesis becomes a system: specialized dense models chained in series, routing external at the task level. Pointed at its own codebase, it starts shipping its own features. It hasn't stopped.

July 2026 — The system governs itself

Review norms become declarative gates; a near-miss becomes a law its agents inherit at birth. Still no orchestrator, no supervisor, no graph. What gets built on Unix primitives are not loops — they are norms.

The stack

Everything below runs local, from the same small case:

RTX 3090 24GB
The inference engine. One card, the whole local compute layer.
Qwopus3.6-27B-v2
Daily driver. Dense Opus distill, MTP speculative decoding, 200K context, vision.
llama-server
Native CUDA build. Serves OpenAI and Anthropic formats to every agent at once.
pi
Lean agent framework — four tools, vision, persistent memory. Pipeline-MoE's substrate.
Pipeline-MoE
The multi-agent room: local and frontier models composed per seat, with review gates.
SearXNG
Private search, self-hosted on the VPS. My queries train nobody's model.
WireGuard mesh
Home ↔ VPS ↔ mobile. The backbone that broke the LAN prison.
CachyOS
Arch-based, scheduler tuned for inference. The OS is part of the stack.

Method

Popperian falsifiability, applied to systems and to my own thinking. If a claim can't be tested, it's not a claim. If a model can't say "I don't know," it's not honest. If an article I publish can't point at a transcript, a commit, or a session file, it doesn't get published.

The local stack isn't privacy as ideology — it's controlling the variables that matter. Local inference makes behavior reproducible; local memory makes the identity scaffold mine to shape; local search keeps my queries out of someone else's training set. The rest — mail, video, backups — stays where it works best. Frontier models get hired, per seat, for the decisions that deserve them. Pragmatism beats purism.

Sovereignty isn't all-or-nothing. It's choosing which dependencies you accept — and which you refuse.

Artifacts

blog.daxzeit.eu
The Pipeline series, architecture forensics, epistemic honesty. Every claim grounded in an artifact.
github.com/DAXZEIT/pipeline-moe
The multi-agent orchestrator, open-source — written entirely by its own agents.
DAXZEIT on HuggingFace
The RA quantization series and model detection work.
Photography
Sony A7II, Samyang 85mm f/1.4, Glimmerglass 3. The one practice here with no agents in it.
Full disclosure, in the house style: this page was drafted by Claude Fable 5 under direction — the same division of labor it describes. The architecture is mine; the typing never was.

Contact

dax@daxzeit.eu