Gradio

SQL distance metrics confirm that CLIP’s shared vector space is geometrically organized. A Gradio UI confirms whether the results are meaningful to a human reviewer. This post builds that prototype and uses it to surface CLIP’s real constraints on a car damage dataset: class confusion in the dent category, complete accuracy inversion on French-language glass shatter queries, and a search engine that returns results for every input regardless of domain fit.