Public Data Site Kit

I built six small websites from Polish public data with my AI agent: housing prices developers must publish by law, smog from state stations, regional statistics, public tenders, a site that crosses four of them, and a weather site that reads sensors inside parcel lockers. All six run the same pipeline. Fetch the data every day, clean it, build one static page per place, and rebuild only when something moved. Most of the work was distrust. A thermometer in a metal box in the sun reads 30°C while the street is at 12, so each city's number is a median, and a sensor far hotter than its neighbours loses its reading. This kit is that pipeline, made generic. You write one JSON config per site: where the data lives, how its columns map, and what you distrust about it. The kit fetches CSV files or JSON APIs, logs every row it drops with the reason, and builds a ranking, a page per place, a search page, a sitemap and structured data, with the source and date under every number. It refuses to build until you fill in a licence check for each source. Standard library Python, no account. The demo builds a sample site from public domain U.S. data, offline.

Built from the pipeline behind six public data sites the author runs. 147 tests, passing on Linux under Python 3.9, 3.11, 3.12, 3.13 and 3.14 and on macOS on Apple silicon under Python 3.9 and 3.14 as of October 30, 2026.