Geometric estimation has long relied on principled models and optimization, but many real-world settings challenge explicit modeling assumptions. In this talk, I will present recent work from my group and collaborators on integrating machine learning into geometric estimation pipelines. The core theme is that learning can complement, rather than replace, classical geometry: for example, it can model noise and uncertainty that are difficult to parameterize analytically, and it can strengthen estimation in ambiguous or underconstrained cases by introducing learned priors about scene structure. I will discuss examples spanning room layout estimation, visual localization, camera pose estimation, structure-from-motion, and image matching, and I will close with a preview of ongoing directions.