spotlight

Why we included this project

Finding the handful of images or audio clips that will actually break your model usually means scrolling through spreadsheets and squinting at thumbnails. Spotlight replaces that grind with an interactive visualization you launch from a dataframe in a few lines of Python, mapping embeddings, predictions, and uncertainty scores onto a canvas where clusters and outliers become visible on their own. ML and engineering teams use it to check where a model is going wrong and catch labeling mistakes, then decide what to curate or re-annotate before another training run. The same call handles images, audio, text, video, time-series, and geometric data, so a mixed dataset doesn't force you to hop between tools. For data-centric work that needs a quick, honest look at what's actually in the data, this is a lightweight companion rather than a heavyweight platform.