A tool for building training labels when there is no ground truth. Open a CSV or Parquet file, describe the thing you are hunting for, and work through the rows it ranks highest. Every judgement refits the model, so the ordering improves as you go.
It is built for a concept rare enough that reading rows at random finds one in a thousand. The whole thing computes in this tab over DuckDB and WebAssembly: no account, no install, and no part of the file is sent anywhere.