Every surface uses the same tokens. The palette shifts from the immersive dark (--im-bg) to classic bone-paper (--classic-bg), but type, spacing, and chrome rules are identical. One bold signature per surface — shader, DotGrid, drop-cap, highlighted word.
A slowly growing archive of what I think about while building. Most pieces are short. A few are long. Cross-posted selectively to Substack and Medium.
On why the outer surface of an AI product should look as authored as its model — and what it costs to keep that promise when the model keeps moving.
A line-by-line tour of the new agent-to-UI spec. What it gets right about render safety. Where it still assumes chat.
Layout confidence. Token-cost ribbon. Attention-head heatmap. Lightweight surfaces that answer the questions users are already asking silently.
Why every AI-shell site looks the same, why it probably shouldn't, and the small set of decisions that change it.
Short. Four rules I keep returning to when letting a model emit UI.
On removing the marketing-shaped object from the top of every page and putting the authored thing there instead.
Most AI interfaces hide their reasoning. A prompt goes in; an answer comes out; the user is asked to trust. That trust is expensive when the system is wrong in small, structural ways — and we are starting to find out how often that is.
I have spent the last year building generative components on top of the A2UI protocol, and the thing I keep returning to is render-time transparency. What the model picked. What it almost picked. What it couldn't see. These are not debugging concerns. They are the UI.
There are three small, lightweight surfaces I've been fitting into AI-native products: a layout confidence overlay, a token-cost ribbon, and an attention-head heatmap. Each earns its place because it answers a question the user is already asking, silently, about the system.
I'm a Staff Software Engineer working at the seam of design and AI. For the last decade I've built generative UI systems, visual interpretability tools, and editorial surfaces that refuse the statistical-average aesthetic of the web.
I write at Design Logic and publish longer research on Medium. My current obsessions are the A2UI protocol, render-safe generative components, and why the shell of a product should look as authored as its model.
Declarative component generation grounded in the A2UI protocol. Render-safe emission, token-cost telemetry, and a small set of components the model can compose without going off-script.
A small set of probes that surface what the model is looking at — down to the attention head. Layout-confidence overlays, token-cost ribbon, and a heatmap that makes "why did it pick that" legible.
Most generative surfaces return freeform markup. That's cheap to ship and impossible to trust. We needed a surface the model couldn't break — and a trail you could audit after the fact.
A 22-component library the model composes against. 99.4% render-safe rate across 40k production emissions. Average per-answer cost down 62% once the token ribbon was visible to end users.
Users trust AI output without a way to see why the model said what it said. Trust without transparency is expensive — and corrodes fast when the system is wrong in small, structural ways.
Three probes shipping in two production products. A Circuits-style thread on attention steering. The heatmap probe lifted self-reported answer trust 31% in user study (n = 84).