Interaction problems in AI and enterprise software, built rather than described.
I'm a product designer working on complex B2B systems — ERP, workflow tooling, AI-enabled platforms. When a design question is about latency, streaming, autonomy or trust, a static comp cannot answer it. So I build the answer first, feel where it breaks, and take it back into Figma once the interaction is settled.
This repo is where those questions get built. One experiment at a time.
The widget reports the window that runs out first, and says so in time — not percent.
A real widget's menu has these items. Their absence is what would give the prototype away.
| # | Experiment | The question it answers | Demo |
|---|---|---|---|
| 02 | Usage widget | When you're spending a finite allowance, what number do you actually need to see? | Live → |
Next.js · TypeScript · Tailwind CSS. Fake data throughout — no backend, no model calls.
Production code. Each experiment is scoped to one question and deliberately stops at the edge of it. The list of things I chose not to build is part of every experiment README.
Denys Sergushkin — Principal Product Designer, 14 years in ERP and AI-enabled B2B software.