Thor Schaeff

Thor Schaeff is a Developer Relations Engineer at Google DeepMind focused on Gemini Audio: audio understanding, speech generation, and the Gemini Live API for realtime voice agents. His work includes teaching developers to build with AI audio and demonstrating how speech generation fits into application infrastructure. He previously worked on developer experience at Stripe, Supabase, and ElevenLabs.[5][8]

Schaeff also uses the name Thorsten Schaeff and the handle thorwebdev. His GitHub profile, displaying the name Thor 雷神 Schaeff, listed “DX at Google DeepMind” as of September 20, 2026.[2][3][4]

On March 5, 2025, Schaeff published a speech-streaming and caching demonstration combining the ElevenLabs API with a Lovable-generated React frontend and Supabase Edge Functions and Storage. The example addresses two practical needs: playing speech while it is still being generated and reusing audio when the same text and voice are requested again.[8]

The request handler hashes the text and voice ID to look for an existing audio file. A cache hit returns stored audio through Supabase’s CDN without generating it again. On a cache miss, the handler requests streaming speech from ElevenLabs and splits the resulting readable stream into two branches: one returns audio immediately to the client, while the other uploads it to storage in the background. Playback can therefore begin without waiting for the storage upload to finish.[8]

At the AI Engineer World’s Fair in June 2025, Schaeff hosted the Voice track and participated in ElevenLabs’ v3 alpha launch. His developer education also included a scheduled August 21, 2025 Shipaton session with RevenueCat’s Charlie Chapman, focused on integrating responsive conversational voice into apps.[1][9]

In Kevin Blanco’s August 3, 2025 account of their DevRelCon conversation, Schaeff emphasized individual, self-serve developers and their progress from trying a platform to shipping with it. He argued that ElevenLabs should bear the cost of giving developers a meaningful opportunity to test AI audio and validate an idea, with starting examples that let them reach an MVP before exhausting credits.[6]

Schaeff described rapid AI releases as a challenge for developer relations: “I think that is inherently our challenge in DevRel now, as everything is moving so quickly, finding that balance of some hype, some groundwork.” His concern was that chasing every announcement could crowd out tutorials and documentation. In the same conversation, he predicted that voice would become part of every app, comparing its spread to responsive design.[6]

Before his voice-focused work at ElevenLabs, Schaeff built customer-facing engineering teams at Stripe across Europe and APAC, then worked in developer experience at Supabase. These earlier roles provide context for his application-building tutorials. In his Voxgig interview, he described learning and teaching as a common thread among his roles.[9][5][3]

His August 29, 2024 Supabase tutorial demonstrated fully local, in-browser semantic search. Schaeff wrote that the idea followed an in-browser Postgres sandbox built in collaboration with the ElectricSQL team using PGlite and pgvector. His subsequent example combined that database tooling with browser-based embedding generation.[7]

The implementation stores text and embeddings locally in IndexedDB through PGlite, generates search-term embeddings with Transformers.js and Supabase/gte-small in a web worker, and uses pgvector for inner-product search. Because the embeddings are normalized, inner product and cosine similarity produce the same query results. Schaeff proposed product search and similar-product recommendations as uses that could avoid a server roundtrip.[7]