This is my working lab. I build AI and software on a deterministic foundation: a math field, z = x·y, where the same input always maps to the same output. The aim is reliability, not spectacle. Systems that are reproducible and inspectable, that do not drift or hallucinate. Everything below is a small, dependency-free demonstration of that idea, free to try in your browser.
A color picker that stores no colors. The saddle z = x·y sits in a wireframe cube; colors are read off the surface by a lens. Drag to turn it, click to pick, or eyedrop any color to find it. ~80 bytes stand in for all 16.7M colors.
Open the picker →A tone picker that stores no audio. A musical scale lives on the saddle z = x·y as bands of equal pitch. Press and move to play; the note is derived from where you are. Pick the scale, nothing stored but the recipe.
Open the tones →Paste or drop any text, code, or directive and shrink it to cut AI tokens, which cost money. The Schwarz Diamond principle: minimal material, maximum meaning. No AI, nothing uploaded, free.
Open the minimizer →Paste any article, post, or idea and harden it: cut the fluff, flag the hedges and unsupported claims, keep only substance. Plus a Harden directive for the AI judgment pass. No AI, free.
Open the hardener →Drop a file or a whole folder and measure how close the code is to the standard: no duplication, just pointers. Finds duplicated literals, cloned blocks, and copies that should be references. No AI, free.
Open the linter →The working library. Data as nested points on z = x·y: the Russian Doll and Car hello-worlds running live, the full API reference, and the AI benefit measured in real tokens. Dependency-free, browser and Node.
Open the API & demos →The heart of it. A whole field of values computed live from one small definition, with no array of values anywhere on the page. Move the dials and every pixel is re-derived. Includes the honest storage-versus-compute number, and where it does not apply.
Open the demonstration →Load a still, zoom into any region, and compare a plain upscale against fractal detail built from the image's own texture. It adds plausible intricacy. It does not recover detail that was lost.
Open the enhancer →The same operation on a thought. A seed idea elaborates self-similarly into a multi-scale tree. Push the vitality dial and watch grounded structure tip into hallucination, the failure mode you usually can't see in an agent.
Open the sandbox →One thing fills the screen. Click it and its parts bloom out. Click a part and it collapses into the new single thing. You drill as deep as the data goes, then it stops. Explore the built-in z=x·y map or paste your own JSON.
Open the navigator →The top-down view of z = x·y. Move the cursor and a position gives one color and one tone, the same every time. The math is the source; color and sound are two renderings of it.
Open the picker →z = x · y is a shape, the saddle, not paper arithmetic. Gather an identity with how it behaves and you get its state, read off the surface by a lens. It is a way of looking, not a programming language. It does not make AI smarter, which is the part that unsettles people; it makes it cheaper and more reliable: send only the slice that is observed, which cuts tokens sharply (and tokens cost money), and a focused, checked context makes a model wander less and stop making things up. It focuses the model, it does not change it. Added detail is labeled synthesis. Revealed detail is real. Where the data ends, the drill-down ends.