Monoalphabetic substitution ciphers, while foundational in cryptography, present significant challenges for manual decryption due to the labor-intensive nature of frequency analysis. This study introduces a novel hybrid approach combining transformer architectures with algorithmic refinement for automated cryptanalysis. We develop a two-stage deep learning pipeline where T5 models first predict substitution patterns from ciphertext, followed by a linguistically-aware correction model. A deterministic resolution algorithm then reconciles conflicting substitutions while preserving positional constraints. The integrated system demonstrates substantial improvements over traditional frequency-based methods, particularly in handling short ciphertexts where statistical approaches falter. By bridging neural pattern recognition with rule-based consistency checks, our method achieves reliable decryption across diverse text lengths and linguistic styles. All models, code, and evaluation frameworks are openly released to support reproducible research in computational cryptography.