Transducer

Inspiration

I wanted Gemini to understand a real-world scene and turn it into something interactive.

What it does

Transducer takes an image or webcam frame, identifies important objects, assigns them semantic software roles, and builds an interactive graph.

How I built it

Gemini returns structured JSON describing the scene. Transducer validates it with Zod, compiles it, and renders a live semantic circuit.

Built with TypeScript, React, TanStack Start, Bun, Google Gen AI SDK, Zod, Tailwind CSS, and Vercel.

Challenges

The main challenge was keeping Gemini's output structured and reliable enough to compile.

What I learned

I learned more about multimodal AI, structured outputs, validation, and using Gemini to interpret physical input.

What's next

I want to add richer inputs, more semantic roles, and connections to real hardware and software.

Built With

• bun
• gemini
• react
• tailwindcss
• tanstack
• typescript
• vercel
• zod