Journal of Digital Information Management Luan Fernandes de Franca, José Everardo Bessa Maia 24 2 2026 https://doi.org/10.6025/jdim/2026/24/2/67-84 https://www.dline.info/fpaper/jdim/v24i2/jdimv24i2_1.pdf This work presents a modular two layer architecture for recognizing isolated sign language video and converting gloss sequences into fluent natural language text. The proposed pipeline comprises: (1) glosslevel sign recognition using spatiotemporal feature extraction (I3D, EfficientNet, MobileNet) with lightweight classifiers, and (2) natural language smoothing via large language models (LLMs) using prompt engineering. Unlike prior claims of continuous translation, this paper explicitly focuses on isolated signs as a practical building block for assistive systems where segmentation is either manually provided or handled by external modules. Comprehensive evaluations on Brazilian Sign Language (Libras) datasets demonstrate high accuracy in isolated sign classification (F1 > 0.97 for I3D-RGB + LR). A critical analysis of near perfect AUC scores reveals dataset limitations that are openly discussed. For gloss to text conversion, we evaluate LLM smoothing on both clean and noise injected gloss sequences and report BLEU scores under realistic conditions. The proposed architecture is scalable, modular, and adaptable to other sign languages, advancing accessibility for deaf and hard of hearing communities.