New tools arrive faster than anyone can vet them. Evaluated tooling means we test AI models, visualizations, and geospatial systems against real public-interest use before we rely on — or recommend — them.

AI & models. We treat models as instruments that need calibration: we sample for accuracy, measure error rates, and document intended use and limits with Model Cards (Mitchell et al.), guided by the NIST AI Risk Management Framework. The caveats go up front, not in a footnote.

Visualization. A chart is an argument, so it has to be honest and readable. We design to web-accessibility standards (WCAG) and use perceptually sound, colorblind-safe palettes (ColorBrewer), avoiding the distortions that mislead.

Geospatial. For mapping we lean on open, auditable GIS — QGIS and PostGIS — and watch for the classic traps: projection choices, normalizing by population, and the modifiable areal unit problem.

What we learn, we publish: the evaluation and the reasoning behind a “use this / avoid that” call, so a newsroom inherits the testing instead of repeating it.

This underwrites our records-at-scale and “make a chart people can read” help-desk work. See our tooling & code or ask the help desk.