A complete design system for an enterprise AI image generation platform — 546 variables, Atomic Design, light/dark theming, built with AI-augmented workflow.

Lumen v1 had no design system — every screen was built ad hoc with hardcoded values. V2 needed enterprise features (auth, roles, preset catalog) that couldn't scale without systematic foundations. No shared design language, no theming infrastructure, no component governance.
The entire system runs on Figma Variables organized into 5 collections: palette (270 raw colors), theme (126 semantic aliases), typography (139 font tokens), radius (7 border values), and responsive (4 breakpoint tokens). Every color traces through a three-level chain: primitive → semantic → component.

Instead of naming colors by appearance, I built a role-based system: brand, accent, neutral, info, positive, negative, warning. Each role has a full scale from 25 to 950 — semantic naming that survives any palette change.

126 theme variables switch simultaneously between dark and light modes. Dark mode first — the natural context for AI image generation where UI shouldn't compete with generated content.


23 molecules and 12 organisms, each following strict rules: reusable elements need 2+ uses before becoming a molecule. Every interactive component has 3 mandatory states: default, hover, click. No unnecessary abstraction.

Token architecture, documentation, and design-to-code mapping — accelerated by Claude Code + Figma MCP. Visual decisions, layout proportions, and interaction design — crafted manually. AI handles the systematic middle; human judgment sets the direction.


8 production screens across 3 layout patterns: dual-panel for generation workflows, full-width for browsing, centered for auth. Every screen assembled entirely from the component library — no one-off elements.
