Blog
A collection of my recent thoughts and works
-
Inside Frankenstein's lab: a full-duplex model at home
I stitched Gemma, Mimi and CSM into a voice model that listens while it talks, on a tiny GPU budget. How it works, what made it work, and what's still open.
-
Intelligence is not instant
Bigger models and more thinking time both kill a phone call. Voice agents have to get the work off the spoken path and keep the conversation moving.