Jarvis
A local voice assistant for the home, inspired by Tony Stark's Jarvis (as any self-respecting assistant should be), wired into my homelab services.
Context
Since language models took off, I get the feeling we don’t talk about privacy as much anymore. I was the first to drop Chrome and Google before ChatGPT came along, in favor of DuckDuckGo and a bunch of other privacy-friendly tools. But with online LLMs, we’ve all kind of resigned ourselves to handing our data to new players.
That’s why I deeply believe in the future of local models, as a way to take back control over our data.
Lately, I’ve been impressed by Mistral’s latest TTS and STT models, and by what Qwen 3.5/3.6 and Gemma 4 can do. It felt like the perfect moment to start a local voice assistant project that could bridge my home services and my daily needs, without ever leaving the house.
How it works
Voice captured and transcribed to text via Mistral’s STT model. The text feeds a ReAct loop on a local LLM, which decides what to do by calling MCP servers: Home Assistant for the house, Kubernetes for my services. The response is synthesized back to voice via Mistral’s TTS model.
Today
In development. The pieces come together as new models drop, the project moves in fits and starts whenever the open source ecosystem gives me something to build on.