The Code Jeweller

Marbella · Independent practice since 2003

Code, cut like a gem.

Artificial intelligence you own. Built here, handed over whole.

The moment

Most people I hand a system to go quiet for a moment. Then they ask how.

The answer is the system itself, running on their own machines, doing work that used to need a person.

The bench

One pair of hands. The specialists behind them.

One name on the commission, one voice accountable, one pair of hands that signs off. Behind them, specialists in every layer, kept in this room and proven on my own work first. Judge them by the work.

Nothing leaves this bench until it earns its place. The work sets the bar; I meet it.

What you receive

Finished. Yours. Still working when I've gone.

I

Code, keys, and title, in your name, on hardware you own.

II

Nothing leaves the room unless you send it.

III

Built to run after I have left.

How I know

I measure before I promise.

Every figure below was taken on my own bench, and each carries its date.

From the bench

Three results that changed how I build.

Most of the bill is not the answer.

91%

On 21 July 2026, each item spent 540 ms reading the same instructions again, and 50 ms answering. That is the ordinary shape of an AI bill. The cut is the 540.

one job · milliseconds per item

Full power, less output.

−21%

Turning everything up made each part faster on its own and the whole a fifth slower. I run the configuration that measures fastest, not the one that sounds fastest.

items per second · n = 30–40

Where more hardware stops paying.

2

On three machines with three different chip designs, throughput stopped rising at the second CPU worker. Knowing that saves buying cores that would do nothing.

items per second by CPU workers · n = 8 per step

This bench, this house, this site — one commission, start to finish. Read how it was cut →

Replayed at the measured rate · 2026-08-05

The bigger model, faster.

Speed follows the parameters that work on each token, not the size on the label. Memory still goes by total size.

16B mixture-of-experts125 tok/s
0
14B dense36 tok/s
0

Pace from runtime counters, 2026-08-05.

One example of scale · 2026-07

A third of a million items. One pass. Nothing left the room.

Estimated~22 h
Actual~10.6 h
0.0 h
0items read and sorted · on hardware bought outright

326,531 classifications. Frontier API at list rates: four figures. Here: €0 to a vendor.

July 2026. Every commission is measured on its own work.

How the work is delivered

Four layers, one hand.

Any one of them is a commission. So is the set.

Intelligence Layer

Systems that decide — not merely answer

Wired into your data and your tools. Faults handled the way any production service handles them. Not a chatbot bolted onto a website.

Hours back, every week
Proprietary

A backend shaped to your business

The layer that is yours alone: data flows, work that acts, the joins between tools you already run. Owned outright. This is most of the actual work.

Leverage no competitor can license
Web & Platform

The interface, backed by something real

Fast, correct front ends wired to APIs and data that hold. What people touch is the surface of a system underneath.

The surface that closes deals
Automation & Systems

Infrastructure that runs unattended

Scheduled work that fails in plain sight and recovers without a person in the room.

The business runs while you don't

Commission

Tell me what needs building.

A few commissions a year, taken on fit. Write with the actual problem, not a brief template; the reply will tell you whether I am the right person for it.