You can't unit test for taste

The author built a data pipeline for their running app "In the Long Run" to populate interactive maps with points of interest along famous routes worldwide, using GeoNames as the primary data source combined with Wikipedia links as a relevance signal. Rather than relying solely on AI, they used an AI coding agent (Claude) in a supporting role alongside traditional data processing tools like Python, Apache Parquet, and DuckDB to filter and process location data. The project highlights how subjective decisions about what constitutes "interesting" landmarks introduce inherent biases into algorithmic curation, which cannot be unit tested.