Many AI medical devices perform perfectly in the lab but fail in the field, for example due to “ungradable” images in the case of retina screening. This talk details the end-to-end engineering journey of a portable, regulatory-cleared retinal camera designed to overcome these deployment barriers through embedded vision. We explore how expensive optics and specialist training were replaced by on-device intelligence and computational imaging—specifically utilizing high-speed sensors and custom illumination and imaging optics to mitigate artifacts such as small pupils, eyelashes and dust. The talk addresses the challenges of enabling novice-led workflows and the fundamental considerations for developing AI models within a regulated medical device framework. We conclude with performance data from real-world primary care deployments, offering a transparent look at bridging the gap between AI concepts and clinical utility.