design-engine

AI-native Website Intelligence and Generation Engine — an AI web design organization that researches real sites, learns design intelligence into graphs, and generates original Next.js websites. Built on the research-engine substrate. A MoreSalamander StudioLabs production.

User Idea → Intent Analysis → Parallel Web Intelligence (6 workers)
→ Design Context Graph → Design Knowledge Graph
→ Design Synthesis → Copywriting → Code Generation
→ Review Agents → Auto-Improve → Build Gate
→ Finished Website → Graph Memory Loop

Originality is enforced, not promised: the site analyzer reduces fetched pages to abstract traits (markup never crosses that boundary), the synthesis model sees only aggregate statistics, and a deterministic novelty gate de-derives any palette that tracks a single source. WCAG AA contrast is math, applied before any CSS is written.

What generation means here

The LLM proposes design tokens, section plans, and copy through schema gates. A typed 16-component library (accessible, responsive, CSS-animated with reduced-motion support) renders the actual code: a complete Next.js 15 / React 19 / TypeScript / Tailwind 4 repository with per-page SEO metadata, sitemap, robots, security headers, and zero raster assets. next build is the QA agent's compile gate.

The agent organization

Research	UX Research · Visual Design · Branding · Competitor Analysis · Industry Research · Frontend Architecture — six workers running in parallel, feeding one Context Graph from live site analysis
Generation	Design Synthesis (gated) · Copywriting (real-brand personas scrubbed deterministically)
Review	Accessibility · Performance · Design · UX/Conversion · Security — deterministic rules over the generated source, then an automatic improvement loop with monotone score checks

The graphs

Context Graph — the current design problem: psychology axes from intent, analyzed sites, extracted traits, section priors. Knowledge Graph — persistent: websites, traits, industries, projects, with count-weighted industry↔trait edges, semantic fingerprint matching, and a scored memory of every generated site (review scores + optional human feedback; no fabricated conversion data). Both emit to a live DataHub as datasets with design-source lineage.

Run it

.venv/bin/python -m design_engine "Create a website for an AI healthcare startup" --build
# then: cd storage_data/sites/<project> && npm install && npm run dev

github.com/MoreSalamander/design-engine · 29 offline deterministic tests · substrate: research-engine · part of MoreSalamander