Case study · 2025 — Present
DesignOS — Local-First AI Design Studio
Chat-driven design systems with multi-format export
Independent builder · Independent
The problem
The problem
Designers and founders bounce between chat tools, Figma, and export hacks. They need AI generation that still respects a design system and ships to real formats — without every action round-tripping through a cloud service.
Outcomes
- Local-first workspace for tokens, components, pages, decks over a hand-rolled SSE chat stream
- Real export paths — ZIP, HTML, PDF (jsPDF/html2canvas), PPTX (pptxgenjs) — plus a Canva hand-off link, not a full Canva integration
- Layered verify gate (static integrity, security, agent, design-library lint/score) chained under one `bun run verify` command
Approach
Conversation revises artifacts; a design-library governance layer (parser/loader/projector/overrides against JSON Schemas) keeps tokens, type, and voice consistent across a project.
Projects live in the browser/workspace as local-first files; WorkOS auth is fully optional, gated behind an env flag and off by default.
Being honest about scope
The project self-describes as "a solid prototype foundation... not yet enterprise-grade" — one bundled reference design system exists so far, and the Canva "export" is currently just a link-out, not an API integration. Worth stating plainly rather than overselling it.
Key tradeoffs
Storage
Chose: Local-first project files
Rejected: Cloud-only multiplayer from day one
Ownership and offline work beat collaborative complexity for v1.
Streaming
Chose: Hand-rolled SSE over a dedicated library
Rejected: A managed streaming SDK
The chat surface is a thin `ReadableStream`/`text/event-stream` endpoint — a library would add weight without solving a real problem yet.
What was hard
- Export fidelity across PPTX/PDF/HTML
What I would change
- Deeper multiplayer later, not as a blocker
- A real Canva API integration if that export path proves valuable
Artifacts
