The large linguistic variations between Tamil and English, coupled with the unavailability of parallel corpora, present enormous challenges to direct translation. Tamil is Subject–Object–Verb, while English is Subject–Verb–Object, which makes direct word-to-word translation problematic. To mitigate this, we suggest an unsupervised Neural Machine Translation with monolingual corpora, back-translation, and deep learning. The model initially translates English to Tamil through synthetic data, and then iteratively improves translations through back-translation. Syntax-aware reordering guarantees grammaticality in spite of structural variations. Our method attains a BLEU score of 0.72, a ROUGE-L score of 0.60, and a METEOR score of 0.62, which proves its effectiveness. The approach improves translation quality with usage and is particularly beneficial for low-resource languages such as Tamil, minimizing large parallel dataset reliance and enabling machine translation in more linguistically diverse areas.