The Vietnamese deaf community faces significant barriers due to the scarcity of accessible communication tools tailored for Vietnamese Sign Language (VSL). This research directly addresses this significant barrier by developing a lightweight, real-time recognition system capable of identifying 10 common VSL phrases using 3D hand key-points extracted from video frames. The system employs a single-layer Long Short-Term Memory (LSTM) model trained on a carefully curated dataset of 1000 real-world videos. Key-points are extracted using MediaPipe Holistic, and predictions are delivered via a Flask API with latency below 300 ms. The proposed method achieved 99.5% accuracy and demonstrating efficient performance on CPU hardware, making it highly deployable in resource-constrained environments such as educational institutions, healthcare facilities, or mobile devices. This work contributes a reproducible pipeline, dynamic frame buffering, and a practical solution for VSL accessibility.