Thor Schaeff is a developer-relations engineer at Google DeepMind working on the Gemini API, Google AI Studio, and conversational audio. His career at Stripe, Supabase, ElevenLabs, and Google has centered on making sophisticated infrastructure practical for developers, from payments and databases to real-time multimodal AI.
Raised near SAP’s headquarters in Germany, Schaeff discovered web development through an SAP summer program and later returned to teach younger students. After working with Google’s London-based Maps team, he joined Stripe in Dublin, where customer-facing engineering and developer advocacy eventually took him to the Asia-Pacific region.
While living in Singapore, he contributed to Supabase before joining the open-source database company. His Supabase writing spans developer education, client libraries, edge functions, AI integrations, and coding assistants. His browser-vector-search project combines Transformers.js, PGlite, and pgvector for in-browser semantic search; another open-source example connects Expo and React Native to OpenAI’s real-time API over WebRTC.
At ElevenLabs, Schaeff focused on multilingual conversational agents and published guides to integrating Gemini and DeepSeek-derived reasoning models with its voice platform. He joined Google DeepMind in November 2025, shifting his attention to Gemini’s audio understanding, speech generation, and Gemini Live API.
- Developer experience for humans and machines. Schaeff argues that developer platforms must work for both people and coding agents. Clear documentation, open interfaces, discoverable examples, and agent-readable instructions increasingly determine which tools AI-assisted programmers adopt. His developer-relations talk connects this argument to his experience at Supabase.
- Audio understanding beyond transcription. His EchoScript example extracts speaker identities, timestamps, languages, emotional cues, translations, and summaries through one structured Gemini request. He also demonstrates prompt-directed synthetic speech, shaping accent, pacing, and character through performance instructions instead of relying exclusively on large voice catalogs.
- Architecture follows operational requirements. Schaeff distinguishes conventional speech-to-text, language-model, and text-to-speech pipelines from native audio voice agents that process conversation more directly. Cascaded systems offer stronger observability, moderation, and intervention; native audio supports more fluid multilingual interaction but introduces practical constraints around debugging, session management, grounding, and reliability.
- Inspectable applications expose real tradeoffs. His Live Jukebox demonstration combines Gemini conversation, tool calling, and Lyria music generation to turn spoken requests into songs. His implementation examples also examine WebSockets, ephemeral credentials, and the latency differences between direct client connections and server-side proxies. A public post about using his voice instead of typing captures his interest in making conversational interfaces useful in everyday work.
