Strot uses a big LLM once to compile a classifier, then runs every decision on a 144M CPU model

You describe a task like ticket routing or text classification, a large model synthesizes the labels and handlers at build time, and a small local model then answers at about 60 ms per record with no LLM calls, abstaining on low-confidence inputs via a calibrated threshold. The author reports 92.5% vs 74.8% on a routing benchmark over naive integration but admits it hits the small model's ceiling on harder tasks like hate-speech detection; MIT licensed, pip install strot-ai (Show HN).