Computational imaging systems depend on the careful co-design of the optical hardware—which determines how light is captured—and the software algorithms—which determines how captured photons are turned into a useful image or estimate. In recent years, machine learning methods have become powerful tools for improving the software side of this design space, achieving impressive gains in speed and accuracy across a range of imaging tasks. However, these methods are not one-size-fits-all; well-designed algorithms must consider the operating constraints of the underlying system, such as its latency requirements, the available data, and the physics of image formation. When learning-based techniques are applied without regard for these factors, it can produce results that are inefficient, difficult to trust, and poorly suited to real-world deployment. In this talk, I will present a set of guiding principles for intelligent imaging system design, demonstrated through challenging applications in synthetic aperture imaging—showing how physics-informed design leads to faster and more reliable solutions.