Hand landmark detection is essential for various applications, including augmented reality, gesture recognition, and human-computer interaction. Various current models are computationally expensive, limiting their usability on devices with limited resources, like embedded systems and mobile phones. We introduce a lightweight hand landmark detection model using a pre-trained MobileNetV2 backbone and a custom CNN-based regression head specifically designed for hand keypoint detection. The model is trained using the CMU-Hand and FreiHand datasets, which provide extensive, diverse annotations of hand poses. Our proposed model achieves high accuracy by reducing the model size and inference time. Experiments show that our model makes an average error of 0.98 mm on the CMU-Hand test dataset and 0.81 mm on the FreiHAND test dataset. Our model also achieves 98.5% AUC and 13.48 ms inference speed.