espinup

An AI-powered 3D room designer that lets you create furniture layouts using natural language

Comment

Inspiration

Interior design is difficult - sketching floor plans considering all dimensions or learning CAD takes forever. I wanted an AI you could just talk to: "create a home office" and it designs the room. With Claude Sonnet 4.5's spatial reasoning abilities together with Strands Agents, I realized I could build an agent that actually understands layouts like a real designer.

What it does

espinup is an AI-powered 3D room designer that works through conversation. Type or speak your ideas ("add a desk and two chairs"), and the AI creates furniture with intelligent placement, proper spacing, and collision detection. You can manually adjust objects or ask the AI for precise changes. One click generates photorealistic renders called "Magic Scene". You can also export as GLB/STL for 3D printing or CAD.

How we built it

AWS cloud-native stack: Amplify (frontend CDN), App Runner (FastAPI backend), Bedrock (Claude 4.5 agent), Transcribe (voice-to-text), S3 (temporary storage). Agentic AI: Built with Strands SDK - Claude has 15+ tools for creating, moving, rotating, and validating furniture. It plans multi-step layouts, detects collisions, and self-corrects. Magic Scene: Dual-AI pipeline - Claude analyzes the scene and writes descriptions, then Gemini 2.5 Flash generates photorealistic renders from screenshots + descriptions. 3D Generation: Parametric models with trimesh/NumPy - all dimensions in millimeters.

Challenges we ran into

WebSocket complexity: Real-time transcription worked but was fragile - switched to simpler REST batch approach. Agent hallucinations: Fixed by maintaining scene graphs, building a resolver agent for collision fixes, and letting the agent visually validate its work. Spatial reasoning: Teaching the agent orientation (chairs face tables) required direction vectors and geometric reasoning.

Accomplishments that we're proud of

  • True agentic AI - Multi-step planning with tool use and self-correction
  • Magic Scene pipeline - Novel dual-AI approach to generate realistic image from 3D design
  • Voice-driven design - Natural speech to 3D layouts
  • Millimeter precision - Real-world accurate dimensions.

What we learned

Using Amplify and App Runner simplifies a lot of deployment aspects. Strands Agents was easier to work with than expected and has all the features expected from an agent framework.

What's next for espinup

  • User accounts and login
  • Cloud save and sharing
  • Real-time collaboration
  • More complex room templates.

Built With

Share this project:

Updates