What it is
An honest rules vs Jev vs LLM comparison in logistics.
What it does: Triages goods-receipt remarks three ways and reports accuracy, latency and cost.
Why it belongs in the showcase: Shows where Jev wins, where it needs criteria tuning or a rule, and the real cost gap.
Caveats: Small dataset; author benchmark.
Evidence
Public MIT repo (1★ at intake) created 2026-10-17, companion code for a Medium article by Mohammed Ahmadi. README test split (100 remarks × 3 repeats): category 41.0% rules / 100% Jev / 100% Haiku; block 66.0% / 72.0% (85.0% with criteria v2) / 98.0%; Jev p50 229 ms vs Haiku 1,523 ms; cost per 100k decisions $2.26 vs $20.31. Jev plus one master-data rule reached 97.0% on block.
Caveats
- Small German-language dataset; read the README limitations
- Block decision still trails the LLM without the extra rule
- Author-reported benchmark
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.