A voice assistant for my house

The Dot runs open-source firmware, and a small server at home does the rest. It isn't all local: after the wake word, a cloud model handles the speech and the answer. The logs, the to-do list and the voice stay in the house.

Guardrails in code

The assistant can change the house to-do list only through a few checked actions, and every change is a git commit, so anything it did can be read back and undone.

The wake word

I trained a custom wake word, “Guist”, on synthetic audio. The first model looked great on my laptop and never woke the Dot: the device turns sound into features a little differently than training did. So I scored the second model on a copy of the Dot's own detection engine before installing it. A laptop score is a promise, not a measurement.

A fix sent upstream

The firmware's Wi-Fi watcher never noticed when my Dot went deaf on the network. I fixed it with an AI agent, tested it on my Dot, and sent it back to the project: HuskerMinion/techo5#85.