Echocardiography is a widely used diagnostic tool for evaluating cardiac function. However, its effectiveness is heavily reliant on the skill and experience of the operator, often requiring significant training. To enhance the application of ultrasound in emergency situations, particularly by non-expert users, we explore the potential of an AI-driven guidance system that can classify specific echocardiographic views. In this study, we introduce a novel framework designed to classify apical echocardiographic views, focusing on the two-chamber (A2C) and four-chamber (A4C) perspectives, utilizing a combination of Variational Autoencoder (VAE) and Multi-Layer Perceptron (MLP). We trained and validated our model on datasets that include publicly available resources (CAMUS and HMC-QU) as well as proprietary echocardiographic images. The VAE was employed to encode 2D ultrasound images into a compact latent space, preserving critical structural features. These encoded representations were then classified as A2C or A4C using an MLP. The model achieved state-of-the-art performance for both apical views and was further validated on three unseen echocardiographic sequences captured experimentally with a clinical ultrasound scanner. In particular, the model achieved an average accuracy respectively of 97% and 93% for the public datasets CAMUS and HMC-QU. This approach has the potential to enhance the accessibility and accuracy of ultrasound imaging in clinical environments, especially in scenarios where expert sonographers may not be readily available, ultimately supporting more effective and timely heart assessments in emergencies or resource-limited settings.