The capacity to automatically interpret a person’s emotional condition from their face has attracted sustained attention in both machine learning research and practical system design. Such interpretation holds particular promise for upgrading the quality of human–computer interaction, enhancing diagnostic support in clinical workflows, and strengthening monitoring pipelines across safety-critical settings. Historically, progress in this area unfolded in two phases: an early period dominated by rule-driven, manually crafted feature representations, followed by a later wave of end-to-end learned models that derive discriminative cues directly from pixel data. This paper traces that arc, examines the relative merits and blind spots of each generation of methods, and introduces a working prototype aimed at the specific task of sensing player frustration during live gameplay. The prototype ingests a continuous webcam stream, isolates and normalises the face region frame by frame, runs a convolutional classifier to assign an emotional label, and converts those labels into a scalar frustration index updated in real time. Alongside the technical contribution, the paper addresses responsible deployment, covering data minimisation, participant consent, and fairness across demographically diverse user groups. Promising directions for extending the system—including the incorporation of audio and physiological channels, lightweight edge deployment, and individualised adaptation—are outlined in the concluding sections.