Can a model that never writes a word make voice AI safer?
A small try-out benchmark: Jev, the open nimble model and STELLA’s own LLM experts, each listening for hidden risk in 26 voice conversations with older adults.
I'm Felix, a PhD candidate at the University of St. Gallen and the University of Zurich, and a voice AI researcher at ETH Zurich's Agentic Systems Lab. I study how people collaborate with AI agents, from making that interaction more natural through voice to organizing and managing teams of agents at scale. On the voice side, I develop STELLA, an open-source platform for deploying voice agents in healthcare, with a focus on vulnerable populations such as people living with dementia. On the organizational side, I explore how agentic systems should be structured and governed, so that people can assign responsibilities, maintain oversight, and trust the agents working on their behalf.
In my spare time, I also enjoy casual photography, capturing the moments and places that inspire me.
A small try-out benchmark: Jev, the open nimble model and STELLA’s own LLM experts, each listening for hidden risk in 26 voice conversations with older adults.
I built a Slack for Claude Code sessions so my agents could work like a team. Then they started solving problems without me, and I had to work out why that bothered me.
The new release of our open-source platform for LLM-based conversational agents cuts time to first audio from about 1.6 seconds to 900 milliseconds.
A critical analysis of lobbying as strategic institutional work, examining how powerful actors construct definitions of rationality through the German energy sector's solar curtailment framework.

