What we do
Topaz builds software that improves itself while it runs. Most programs are frozen the moment they ship — they do exactly what they were written to do until a person rewrites them. A Topaz program treats its own behavior as adjustable and tunes it continuously from live signal, the way a model trains, except it happens in production and on its own.
Why now
Until now, software that adapts meant a research team, a training pipeline, and months of work — so almost nothing you use actually does it. Topaz makes adaptive software something you write in one language and ship, instead of a project only a lab can afford.
That is the shift: continuous learning becomes a default property of software, not a specialist feature.
The first product
A landing page that optimizes itself. A/B testing needs roughly 27,000 visitors to call a single winner — traffic most sites never get, so their pages just sit there. A Topaz page learns from the traffic it already has.
Here is how. The page is compiled into a differentiable program, so every visual choice — layout, spacing, color, emphasis — is a continuous parameter rather than a fixed value. Visitor behavior (scroll depth, hovers, dwell time) is a dense signal, far richer than convert-or-not, and the page follows it the way a neural network follows a gradient. It does not pick the best of a few human-written variants — it searches a continuous space of designs and settles on ones no one drew by hand.
The stack underneath
The product runs on a stack we built from the ground up. Each layer is shipped and working today.
- Sutra, our language. Logical operations over vector spaces compile at build time into differentiable tensor graphs, so an entire program is one continuous, trainable function. Published on PyPI as
sutra-dev(v1.1.0), with a NeurIPS 2026 submission. - An operating system written in Sutra, GPU-native and neuro-symbolic, where the whole running system is a single inspectable network. Working prototype with a terminal, calculator, and click-driven GUI.
- Self-optimizing pages, the first product — the same stack at ground level, where the page itself learns from every visitor.
Who's building it
Topaz is built by Emma Leonhart, a self-taught deep-tech founder who shipped the language, the research paper, and the OS prototype solo, without waiting for permission.
Request early access
Topaz is rolling out to early users in waves. Join the list to be in the first — and to follow the self-optimizing pages product as it ships.
Rolled out in waves. No spam, no list-selling.