Stellar infrared excesses can indicate various phenomena of interest, from protoplanetary disks to debris disks, or (more speculatively) techno-signatures along the lines of Dyson spheres. In this paper, we conduct a large search for “extreme” infrared excesses, designed as a data-driven contextual anomaly detection pipeline. We focus our search on FGK stars close to the main sequence to favor nonyoung host stars. We look for excess in the mid-infrared, unlocking a large sample to search in while favoring extreme IR excess akin to the ones produced by extreme debris disks (EDDs) and/or planetary collision events. We combine observations from ESA Gaia Data Release 3, the Two Micron All-Sky Survey, and the unWISE version of NASA’s Wide-field Infrared Survey Explorer (WISE), and create a catalog of 4,898,812 stars with G < 16 mag. We consider a star to have an excess if it is substantially brighter in the W1 and W2 bands than what is predicted from an ensemble of machine learning models trained on the data, taking optical and near-infrared information as input features. We apply a set of additional cuts (derived from the machine learning models and the objects’ astronomical features) to avoid false positives and identify a set of 53 objects, including one previously identified EDD candidate. The typical infrared-excess fractional luminosities we find are in the range 0.005–0.1, consistent with previous EDD candidates and potential planetary collision events.