Decision engine

The easy calls, made in microseconds.

A small model reads every inbound message and routes it: answer from a template, hold for a person, or hand to a larger model when it is not sure. It runs on a CPU, with no GPU and no model host, and legal, medical and irreversible requests always go to a person.

94.9%sorted into the right intent, on messages it had never seen
100%legal, medical and irreversible requests held for a person
~33 µstime per decision, on a CPU