Human activity recognition (HAR) and indoor localization are essential components of intelligent in-home systems, particularly for supporting the safety and independence of elderly individuals living alone. Vision-based methods are constrained by lighting conditions, require an unobstructed line of sight, and raise significant privacy concerns, while wearable systems depend heavily on user compliance and are difficult to maintain for continuous, long-term monitoring. In this work, we propose a real-time, interpretable method based on Wi-Fi Channel State Information (CSI) to jointly perform activity recognition and localization without the need for visual input or body-worn sensors. The model integrates an attention-based encoder to extract key CSI features and employs a wavelet-transform-based Kolmogorov–Arnold Network (KAN) to capture multi-resolution motion patterns and nonlinear spatial-temporal relationships. Our model achieves 94.86% accuracy in activity recognition and 98.92% accuracy in localization on the JARIL dataset, outperforming existing baselines. This framework holds promise for privacy-preserving and unobtrusive health monitoring applications in smart home environments.