A one-person laboratory.

For the clients I serve — organizations whose data cannot leave the building — this is an advantage, and I treat it as one. A single point of contact, no subcontracting, no data routed through a third party for internal-org reasons. You know exactly who has access to what: one person.

Creolizing AI

Karaibart's Caribbean anchoring is cultural and methodological, not geographical. There is no promise of a Caribbean address, and no folklore. The anchor is creolization, in the sense Édouard Glissant gave the word: a process where what arrives from elsewhere is neither rejected nor endured, but appropriated and recomposed into something unpredictable and autonomous.

Applied to AI, this is not a decorative metaphor — it is an exact description of what this ecosystem does. Open weights from China. A runtime from Cupertino. Libraries from everywhere. Assembled on a machine that owes nothing to anyone, under a memory budget you control, with rules you wrote yourself.

Take the world's models, and run them at home.

The name itself predates all of it. Karaïb and art — chosen for an entirely different project, years before I worked on AI. Nobody had noticed that the fourth and fifth letters already spelled AI. KarAIbart. Which is precisely what creolization does: it recomposes without warning, and the meaning arrives afterwards.

It is also an ethic of constraint. You do not have a thirty-GPU cluster; you have one machine. That limit is not a handicap to route around by renting compute elsewhere — it is what forces the right decisions. A single memory budget forces you to know which model deserves to be resident. The absence of cloud forces you to measure what you actually consume. Resourcefulness under constraint produces more honest systems than abundance does.

How I got here

Proof, not promise

The method comes down to eight rules, and each one comes from a real incident rather than a textbook.

• Hexagonal architecture — decision logic is pure, side effects live at the boundary. You test an eviction policy without starting a process.
• Dependency injection — a package extracted through Protocols, three heterogeneous consumers. Proof the abstraction is right comes from the second consumer, not from the tests.
• Strangler fig — extraction behind shims, one consumer at a time, never a big-bang. Validated by roughly 1,600 tests at every step.
• Test-as-you-go — full suite and quality gate after each sub-batch, not at the end.
• Adversarial review — a multi-agent workflow before the gateway's first commit: 10 confirmed issues out of 22 raised. The false-positive rate is part of the result.
• Measure-driven decisions — phys_footprint rather than ps rss, cos = 1.0000 verified before migrating 1.7 million vectors, tok/s measured before any adoption.
• Security by default — enforced loopback, anti-OOM cap, anti-DoS guards, telemetry off. Not at the end of the project.
• Honest accounting — never claim an unverified state. An OOM taught me the gateway could believe it had room when the system knew otherwise.

The limits, stated up front

Naming what you do not do costs a few engagements and wins better ones.

• I do not deploy on large-scale NVIDIA GPU clusters. My specialty is Apple Silicon unified memory and single-machine inference, not multi-node orchestration.
• I do not do pre-training or large-scale fine-tuning. I assemble, optimize and harden open-weight models.
• I do not take engagements where data must go to a third-party cloud. That would contradict me, and you would be right to stop believing the rest.
• I am one person. An engagement needing five people in parallel is not for me — and I will say so on the first call rather than in the third month.

This site respects what it sells

No downloaded fonts. No third-party scripts. No trackers, no cookies, no consent banner — because there is nothing to consent to. The HTML is pre-rendered by a few-hundred-line generator written on the Python standard library, with zero dependencies. The figures shown are regenerated by a measurement command, and that command is published next to the number.

Open the developer tools, Network tab. Count the domains.

That is the one proof that cannot be plagiarized.