We propose a novel approach for American Sign Language (ASL) recognition that combines Google MediaPipe's real-time hand and body landmark tracking with a lightweight deep learning model trained on the MS-ASL dataset. We called this model “Key Frame MLP”, and it extracts the key frames features from the sequence of hand and pose landmarks, enabling efficient recognition without requiring RGB input. Evaluated on the MS-ASL 1000 dataset, our approach achieves 61% top-1 average class accuracy, demonstrating strong performance relative to its simplicity. The entire system is optimized for real-time operation and low computational cost, making it suitable for deployment on edge devices. These results highlight the effectiveness of combining modular MLPs with fast landmark-based inputs for scalable sign language recognition.