This study presents an experimental implementation of a Transformer-based translation system inspired by Vaswani et al., trained on the Samanantar English–Indic corpus to evaluate multilingual performance across 11 languages. It details modular PyTorch components for custom tokenization, positional encoding, multi-head attention, and feed-forward networks, alongside a training regimen with warmup scheduling and label smoothing. An accompanying Streamlit prototype enables interactive validation of translation quality and scalability, bridging theoretical foundations with practical deployment.