A web tool for labeling land cover in Seattle from a small user-annotated collection of examples, built on precomputed embeddings from Ai2's OlmoEarth model. 'Painting' a few cells as water, canopy, or other generated category classifies the remaining 16,000+ cells by nearest match. Cells that don't clearly fit a class are left unlabeled rather than forced into the nearest guess, and cells where two classes nearly tie are flagged as uncertain. The tool is client-side, with no model calls at runtime.

This project doubles as a small study in evaluating a foundation model through interaction design, and in directing Claude Code across a full pipeline from computing embeddings offline, diagnosing artifacts in the model's raw output, to shaping the interface around what the data supports.