Strot releases an open-source decision compiler that runs local models on records
Original titleShow HN: Strot – Compile an LLM's intent once, run decisions locally
Strot is a Python library that uses an LLM once at compile time to build a frozen artifact, which then routes or classifies each record with a local decision model.
The post reports 92.5% grounded routing accuracy on 398 records with the 144M-parameter Julia-1 model, compared with 74.8% for a naive baseline, at about 60 ms per record and no marginal LLM cost.
The library is MIT-licensed and installs via pip install strot-ai.
The post explains a compile-once design where an LLM is used only to build a frozen local decision artifact, which is a useful pattern for teams weighing per-record LLM costs.
Source: Hacker News · Show HN, AI (20+ points) · github.comPublished · added here