Voice-first Reverse Search
Voice agent for finding things you can't fully describe, like a half-remembered song, book, place or product. It listens, draws controls on screen as the conversation goes, and searches the source that fits.
Architecture
Two-way audio over the Gemini Live API with server-side voice activity detection and barge-in, through a Bun proxy that issues ephemeral tokens so the API key stays on the server. The agent generates its interface with A2UI and routes each query to a domain search tool (music, books, maps, shopping, or parallel web search) with caching and rate limiting. A shader-driven 3D presence reacts to the voice. Deployed on Cloud Run with Terraform, Secret Manager and a budget alert.
Highlights
- Built at a hackathon, deployed live
- Interruptible real-time voice
- Agent-generated UI
- Live tool tests and LLM-as-judge evals
- Earlier version benchmarked models on latency and cost