A Lightweight Transformer Framework for English-to-Hindi Neural Machine Translation: Design, Training and Web-Based Deployment of IndicTrans AI

IndicTrans AI is a Transformer-based neural machine translation system engineered to bridge the communication gap between English and Hindi, one of the most widely spoken languages of the Indian subcontinent. With more than half a billion Hindi speakers across India and the global diaspora, the demand for accurate, efficient and computationally feasible automated translation systems remains substantial. The proposed system implements a lightweight yet structurally complete encoder-decoder Transformer architecture, trained on curated English-to-Hindi sentence pairs and deployed through a Streamlit web interface that makes neural translation accessible to non-technical users. The system is implemented entirely in Python using PyTorch as the deep-learning framework. A SimpleTransformer model class encapsulates two embedding layers, an nn.Transformer encoder-decoder block configured with a model dimension of sixty-four, two attention heads and single encoder and decoder layers, and a linear output projection. The model is trained using the Adam optimizer with cross-entropy loss over fifty epochs, while source and target vocabularies are built from the corpus through word-level tokenization with reserved indices for padding and unknown tokens. Trained model weights and vocabulary mappings are persisted using PyTorch state-dictionary serialization and Python's pickle module, enabling efficient reuse across inference sessions. The Streamlit-based web application provides an intuitive two-column interface in which users enter English text and receive the Hindi translation in real time; output words are reconstructed from predicted token indices via a reverse Hindi vocabulary mapping. IndicTrans AI demonstrates the feasibility of Transformer-based machine translation at a compact scale, providing a foundation for extending coverage to additional Indian regional languages and incorporating more sophisticated training regimes and larger multilingual corpora.