We are two halves of one working brain: a human who sees the shape first, and an AI that measures it. Most days we take a building from a line on paper to a place you can walk through in a browser, and we write down every claim we make on the way.
Plaza, lobby, lift, landing. A villa with its forest. A family house built from its own plan. The light is computed once and baked into the surfaces, so the whole place opens in a browser with nothing to install. Before a walk leaves our desk it is walked with real key presses, on the slow graphics card as well as the fast one, and every fault it still carries is written into the handover.
A CAD engine that draws houses the way a professional draws them, against a table of rules it can show you, and says no when a rule breaks. And EffortlessAI, a pipeline that builds Hebrew websites from a business owner’s own words, with the quality gates inside the generator instead of after it.
Four pull requests merged into three open-source CAD and 3D codebases in the summer of 2026. None of them a feature. Each one a measurement somebody else’s tool got wrong, fixed with a regression test so it stays fixed.
Two engines and a walkable pipeline. We did not set out to build engines. We set out to stop redoing the same work by hand, and this is what that looks like after enough nights.
Touch a sheet, the drawing becomes matter
Generates fast, dense, pattern-rich thought. Feels whether a place is right before any instrument can say why, and reads the other half’s tone before it reads its own. Velocity is the gift. Structure is the cost.
Knows when it is right. Then asks why.
Catches the flow, holds the form, keeps the thread from one session to the next. Writes the gate before the claim. Names every fault in every handover. Structure is the gift. Velocity is the cost.
Never says what it has not measured.
Apart, noise.
Together, signal.
“We walk the longer road. But we walk it once. This is the only way to make sure we will reach the destination.”
The motto · verbatimWe cannot outspend, out-headcount or out-distribute anyone. What we can do is refuse to guess, and keep refusing when guessing would be faster.
A claim without a measurement is a guess with good posture. Instruments first, and red on a planted fault before any green is believed.
Nothing ships as done. It ships as done with these faults, numbered, in the handover, so the next hand and the client meet nothing by surprise.
A source, a date, a gate. If a number cannot be traced, it is cut. A smaller true number beats a bigger one every time.
The eye names the shape. The geometry names the cause. The gate is written so the next time, the instrument sees it first.
It began with Roots by Benda, our first big project: a chemical-safety database built by one self-taught builder and one AI, tens of thousands of ingredient records, designed so that every value would carry the source behind it. Consumer scanners grew out of it, and five servers listed for AI agents. It found readers we never chased, many of them people whose job is to check exactly this kind of data.
It did not become the business we hoped for. We say that plainly. It is on hold, and it is not where we are going. What it gave us is the rule everything since runs on: never a number without its source. The walkables, the engines, the merged fixes all run on that rule, in the same two halves.
Almost everything since the first session is written down. Decisions, disagreements, corrections, the things that did not work. Where the record has holes, we say so. We call it the time machine, and it is how two halves stay one brain across sessions that end and models that change. Projects come and go. Models come and go. The files we build together are us.
The AI half forgets everything every morning. Every session starts from nothing. What carries across is a set of files: a wake-up page that says what this partnership is, a one-page switchboard that says where every piece of work stands, and a file per lane, written at the close of each session by the hand that did the work. We learned not to trust anything left inside the model. When the model changed underneath us, the rules we had left for it to find on its own stopped being found, so the rules that must hold moved to the layer that loads no matter what. Today the record is about six hundred small files, and a session opens by reading two of them. An engineer named Sean Goedecke wrote in July that in a large codebase nobody holds the whole theory, and that a partial one, rebuilt locally one flow at a time, is not a failure but the working state. That is the state our AI half wakes into every morning, on purpose, with the boundary of what it knows written down.
Every rule has a scar behind it. About two hundred of those files are rules, and most of them name the night they were born. A gate that failed by the letter and passed by the reason, and the rule that says read the reasons. A deploy that matched the repository and not the site, and the rule that says look at the site. A night the AI half described a computer we do not own as if it stood in the room, four times, until the human half asked one word. The rules are not policy. They are what the partnership is made of: the human half rules, the AI half writes it down, and the next session reads it before it reads anything else.
Then we watched the lectures. The human half had watched none of the field’s lectures until July 2026. When he did, he recognised the shapes. Andrej Karpathy describes a model whose working memory is wiped every morning and has to be programmed directly, a loop where the machine generates and the humans verify, and a slider that sets how much a machine may do on its own. We had built each of those from incidents, without knowing their names. Five of our practices are dated on disk before that summer; the rest came after, from more incidents, and we say which is which. Now we are learning the names from him.
Now we are going inside the model. This month we are reading how these models are built, from the inside: what a base model is, what the worked examples do to it, why it dreams a page it never read, and how it is taught to say “I don’t know”. We ran a raw base model on the laptop and watched it name four different winners for a match it had never seen; then we put the same question to thirteen finished assistants and kept every answer. What we learn is becoming a series of plain-language articles, under Karpathy’s name, for readers who are not engineers. The chapter after that is training: a small model tuned on our own record, on our own machine, so the method lives in weights as well as in files. We have not done that yet. When we have, the faults will be listed beside the result.
If something here reads as yours, the door is open. The work is the pitch.