Transformer from scratch

Built the Transformer architecture entirely from scratch, including Multi-Head Attention, positional encoding, encoder-decoder attention, feed-forward networks, masking, residual connections, layer normalization, and a custom Byte Pair Encoding (BPE) tokenizer. The project also implements greedy autoregressive decoding and an end-to-end training pipeline for English-German machine translation.