Communication accessibility for Indonesia’s deaf and hard-of-hearing community remains severely limited. Currently, over 2.7 million individuals rely on the Indonesian Sign Language System (SIBI), yet they frequently encounter communication barriers due to the lack of widely accessible translation tools. Existing solutions often depend on costly wearable devices or camera-based systems restricted to static gesture recognition. To bridge this gap, this study proposes a comprehensive SIBI recognition system that integrates Bidirectional Long Short-Term Memory (Bi-LSTM) networks with MediaPipe Holistic feature extraction. Deployed as a fully offline PyQt5 desktop application, the system ensures privacy, low latency, and independence from internet connectivity. A custom dataset comprising 10 distinct gesture classes was collected and processed into 1,662-dimensional landmark vectors over 30-frame sequences. The underlying model is a three- layer stacked Bi-LSTM configured with L2 regularization and an Adam optimizer. Evaluation results demonstrated a robust validation accuracy of 97.00% and an F1-score of 0.97, with 9 out of 10 gesture classes achieving a 100% real-time recognition success rate. Furthermore, the application exhibited excellent resource efficiency, running at approximately 26 FPS, consuming only ~30% CPU and ~280 MB RAM, with an end-to-end latency of ~230 ms. A stress test confirming stable operation for up to 50 minutes validates the system's viability as a practical, offline assistive technology tool for educational and public settings.