This project aims to build a lightweight, end-to-end multimodal model for emotion recognition using only audio and visual input, and see how far we can push its performance under real-world constraints.