신하니의 실험 노트

Field notes

The archive follows a repeatable pattern: identify an operational bottleneck, build a workflow around rules or structured AI use, then record the measured change and remaining failure modes. Examples include a 100-address geocoding set with an 88% automatic-classification rate and a cosmetics review test built from 996 labeled sentences. The author also records validation limits, such as one current user for the address service and a 119-sentence holdout used to test reproducibility.