DeepLens Engine for Focus: A Multi-Layered Framework for Real-Time Attention Monitoring in Productivity Environments

The proliferation of digital distractions has made sustained attention increasingly difficult to maintain. Studies indicate that average screen attention spans have declined from 2.5 minutes in 2004 to approximately 47 seconds in recent years, with interruption recovery requiring over 20 minutes on average. While artificial intelligence-based attention monitoring systems have emerged as potentialsolutions, existing approaches suffer from three fundamental limitations: lack of standardized attention state definitions, reliance on single-modal detection methods, and absence of temporal context leading to excessive false positives. This paper presents the DeepLens Engine for Focus (DLEF), a novel multi-layered framework addressing these limitations through systematic integration of deep learning classification, auxiliary validation, and temporal heuristic smoothing. Grounded in productivity science—synthesizing insights from deep work theory, deliberate practice research, flow psychology, and execution frameworks—we establish formal definitions for six attention states: Focused, Phone, Absent, Drowsy, Looking Away, and Bad Posture, operationalized through observable physical indicators. The DLEF architecture comprises four synergistic layers: (1) a primary YOLOv11n-cls classifier trained on approximately 3,000 augmented images derived from 1,080 base samples, (2) auxiliary validation modules utilizing MediaPipe for gaze estimation and eye aspect ratio computation, (3) a temporal heuristic layer implementing the novel 20/30-second rule for sustained distraction confirmation, and (4) a context-aware response layer for non-intrusive intervention. Experimental evaluation on 210 held-out validation images demonstrates 94.76% classification accuracy, with 100% precision on critical classes including Phone and Absent. The temporal heuristic layer substantially reduces false positives compared to frame-by-frame classification while maintaining high sensitivity. Real-time performance achieves 15–18 FPS on consumer hardware. We implement thecomplete framework as DeepWork AI, a privacy-preserving web application with Pomodoro-style session management and comprehensive analytics.