Attributed to the omnipresence of the radio signals for communications, sensing and recognition utilizing the Wi-Fi signals has significant advantage in terms of accessibility over conventional sensing means such as the camera. However, utilizing the raw Wi-Fi signals to capture in-air handwritten signatures for identity verification is yet a challenging task. In this paper, we propose a system for identity verification based on the handwritten signature signals captured by the Wi-Fi Channel State Information (CSI). A triplet network is adopted to learn the correlation between the captured signals and the user identities. To facilitate a fast converging loss model, a kernel and the range space learning is initially adopted for mining the triplet inputs. Subsequently, the triplet network is trained on a ConvNet structure based on the mined triplet inputs. Our experiments on a Wi-Fi dataset collected in-house show encouraging verification accuracy with faster training loss convergence comparing with that of the baseline triplet network and the Siamese network.