Wi-Fi sensing—especially human activity recognition (HAR)—is emerging as a device-free means of understanding human–space interactions for smart healthcare and elder care. However, most learning-centric HAR systems focus on improving feature-extraction efficiency or advancing neural-network architectures while overlooking physical information—such as location—which is crucial for accurate activity recognition and interpretation. To close this gap, we introduceDWSen, a Dual-Path Wavelet-Attention Kolmogorov-Arnold Network (KAN) framework for joint activity and indoor location sensing. Specifically, DWSen is a compact multi-task model that combines two complementary encoders with a lightweight decoder. First, an Attention-Masked ResNet-1D (AMRN) suppresses noisy subcarriers and thus highlights discriminative temporal patterns. Meanwhile, a Dual Laplace Wavelet Convolution and bidirectional GRU (Bi-GRU) branch (DLWCB) injects physics-aware, band-pass time–frequency cues and captures long-range temporal dependencies in CSI sequences, thereby enhancing the accuracy and robustness of joint activity and indoor location recognition. Thereafter, the dual features are fused at a late stage and decoded by Wavelet-based Kolmogorov-Arnold Network (WavKAN) heads, which incorporate analytical wavelet kernels as priors and combine them with a nonlinear path and batch normalization to generate activity and location logits. Moreover, training employs an uncertainty-based Adaptive Weighted Loss (AWL) that dynamically balances the objectives, thus mitigating negative transfer without manual tuning. Additionally, we collected an activity- and location-sensing dataset, ElderAL-CSI, covering six activities at nine locations performed by three participants. Extensive experiments on ARIL, CSIDA, and ElderAL-CSI demonstrate that our method attains 94.24%/98.20%, 94.37%/99.65%, and 99.56%/85.20% accuracy (activity/location), respectively, and offers competitive or superior performance compared with recent baselines across most settings. These results point to a low-cost, privacy-friendly path toward reliable ambient monitoring in real homes.