In this paper, I describe a Bangla Sign Language (BdSL) translator, which identifies 60 dynamic signs with the help of landmark-based machine learning methods. The purpose of the system is to eliminate the communications barrier between the Deaf and Hard-of-Hearing population and the hearing population in Bangladesh by translating the sign language to text. A multi-stage preprocessing pipeline is trained which includes MediaPipe Holistic landmark extraction, coordinate normalization and temporal quantization which convert the gesture videos into a format of structured sequential feature representations. Four methods of classification are considered, and they are LSTM-based temporal models and conventional machine learning models. The experimental findings have indicated that temporal modeling has a great contribution to recognition performance. The LSTM-D5MAX set has the best performance with the highest accuracy and good discriminative properties of 76.90 percent, and the performances of DTW-SVM and regular LSTM are moderate. In its turn, the classical SVM demonstrates low efficiency as it is incapable of capturing temporal dependencies. The proposed system guarantees the privacy of users since it does not make use of raw video frames but only skeletal landmark data which is computationally efficient enough to be used in real-time mode. This work provides a scalable base to more vocabulary BdSL translation systems and leads to privacy-protective intelligent assistive communication technologies.