Luminova
You describe. AI designs. Everything editable.
6 followers
You describe. AI designs. Everything editable.
6 followers
Luminova is a design editor where AI is your co-designer, not just a chatbot on the side. Type natural language into the chat "add a blue rectangle" and watch changes happen live on canvas. Drop in a screenshot and say "recreate this." The AI sees your canvas, understands what's selected, and executes multi-step designs autonomously. Three AI modes (Chat, Plan, Agent) scale from quick tweaks to full layout builds. Use @-mentions to reference objects and /workflows to save reusable styles.
I started building this 10 months ago as an undergrad. The idea was simple, what if designing felt like talking to a teammate? You describe what you want, and an AI builds it on an editable canvas while you watch and collaborate with it.
I had a working v1 months ago but kept polishing instead of shipping. Today I'm finally shipping it. I wanted to build what doesn't exist yet, a design editor where you can @mention any object on canvas, where the AI plans before it builds, self-corrects its own work, and treats every element as an editable layer. Not generate-and-hope. Actual collaborative design.
Some things I'm proud of:
The @mention system, I took inspiration from ByteDance's SeedDance 2.0 and thought, why can't I click any canvas object and reference it directly in chat? So I built it. "Make @headline bigger" just works.
Plan mode, the AI doesn't just execute blindly. It thinks through the design, asks you interactive questions (vibe, palette, fonts), then builds after you approve.
Agent mode, the AI iterates autonomously, takes screenshots of its own work, self-corrects, and keeps going until it's right.
Reference image recreation, screenshot any design you like, paste it in, say "recreate this." The AI analyzes colors, layout, and typography, then rebuilds it as editable layers.
I rebuilt the architecture multiple times, started with REST APIs, moved to WebSockets for real-time streaming, rewrote the canvas state system. Every painful rebuild made the experience better.
Built this solo, tested with design students and professors. Would genuinely love your feedback, this is the launch I kept putting off, and I'm finally shipping it.
RiteKit Company Logo API
@loveneet_singh "Not generate-and-hope. Actual collaborative design." That's the framing that makes this click. Two questions: (1) When the AI self-corrects in agent mode, does it ever get stuck in a loop over a subjective choice (e.g., shade of blue)? (2) How do you handle the user overriding a correction mid-stream without breaking the agent's context?
Hi @osakasaul thanks for your comment!
The AI in agent mode is guided more towards action, it's designed to lock in the vision through plan mode beforehand, so it doesn't get stuck while creating designs. In agent mode specifically, the AI is instructed to bias towards action it makes a quick decision by selecting a design choice and moves forward with it rather than looping over subjective options.
I've implemented a barge-in pattern similar to Cursor IDE, where any new request enters a queue that can be interrupted easily so the agent picks up the latest request. This is handled using an interrupt pattern in LangChain.