The ambitious goal of the COMPASS project is to revolutionize computational sensing by going beyond the conventional "sense first, analyze later" paradigm. The project's primary goal is to recover and improve high-resolution signals from huge, flawed datasets, particularly those collected from Earth Observation (EO) satellites, which are frequently noisy, fragmentary, and devoid of ground-truth references. In order to accomplish this, COMPASS creates a potent end-to-end modeling framework by fusing the advantages of learning-based optimization with generative and discriminative machine learning. This method guarantees that the recovered data are actually helpful for additional analysis, such as categorizing land-cover types and identifying environmental changes or anomalies over time, in addition to enhancing the quality and dependability of signal recovery. COMPASS strives to make scientific data analysis more reliable, transparent, and ultimately more influential for comprehending our world by quantifying uncertainty and spotting potential model hallucinations.