There is an increasing need for sophisticated neural machine translation systems due to the need for trustworthy worldwide data transfer. The opaque black-box behaviour of large pre-trained models limits customisation and interpretability, despite their impressive performance. To tackle this issue, we develop a transparent, from-scratch Transformer implementation for translating from English to Hindi using PyTorch. As part of a whole encoder-decoder architecture, the concept’s main components are multi-head self-attention and sinusoidal positional encodings. Its competitive BLEU scores after training on the cfilt/iitb English-Hindi corpus demonstrate the feasibility of a custom-built approach. It is a helpful place to start for further study and comprehension of contemporary NMT structures.