Hackathon · Jul 2026
Ghar
Talk to a live AI interior designer. She designs your home room by room, walks you through it and hands you a build pack for your architect.
Deployed website ↗Examples ↗Video ↗Code ↗
- EventGoogle DeepMind hackathon, Bangalore
- TrackLive API, real-time multimodal
- TeamWith Abhishek
Problem
Build a real-time multimodal app on the Gemini Live API in five hours. A live session with audio and video lasts two minutes, and the camera sends one frame a second.
Designing a home usually means describing it to someone, waiting for drawings and going back and forth. We wanted that to happen in one conversation, with no prompt box.
What we built
- A voice session with Asha, the designer, who calls seven tools by herself during the conversation: imagine a room, refine it, play a tour, compile the brief.
- Three ways in: describe your home, upload a photo of a room to fix, or show it on the camera.
- A generation server that renders four room concepts in parallel, room tours as video, and a concept floor plan.
- A build pack with budget, materials and approvals, grounded in Google Search with citations.
Design choices
- Tools reply at once and do their work in the background, so the conversation never stalls.
- The full image model for edits: the faster one was twice as quick but drifted from the room's look.
- Tour videos take 40 to 80 seconds to render, so the tours in the demo are pre-rendered.
Result
Four concepts appear in 13 to 18 seconds. I built the live session, the designer's persona, the tool dispatcher and the concept rail; Abhishek built the generation server, tours and build pack.