Relocalization, the task of positioning a new image within a set of images of a scene with known poses, is a foundational challenge of computer vision. Where most SotA advances focus on accuracy and robustness, little progress has been made in relocalization systems for lightweight hardware. Taking advantage of recent advances in efficient NeRF rendering, we demonstrate how a combination of massive sampling and simple filtering can be used to train a lightweight relocalization model to be both significantly more accurate than competitively lightweight works, AND more robust to extreme viewpoint changes, all while remaining above ~30 FPS on portable hardware like the Nvidia Jetson Orin. We hope our research paves the way for further investigations into smaller models and more data for various computer vision tasks.