Efficiency Lab

Small models for Swiss SME invoice-line classification. Same 74 held-out-vendor test lines, Apple Silicon laptop, MLX.

Accuracy, speed and memory
Model	Accuracy		Median ms/line	Peak GB

Synthetic data. 74 test rows means about +/-6 points of uncertainty. Latency varies by about 20 ms between runs. Code: github.com/ChadSaglam/efficiency-lab