Rust, solo, Lucerna Labs. I build a runtime called Ordo one process, no webviews, no subprocesses, everything moves over a typed bus. People tell me that's too strict. It is, on purpose. Came to ML from RF and codec work, and the thing that still surprises me from my own research 48 targeted "scar" memories beat 240 bulk ones for patching a compressed small model. Correct me when I overstate something. I mean it.