The process of retrosynthetic analysis, introduced by Corey, systematically deconstructs complex molecules into simpler precursors, providing a logical pathway for chemical synthesis. Here, we propose an innovative AI-driven retrosynthesis framework for drug discovery leveraging Large Language Models (LLMs) and advanced computational tools. This "retro drug discovery" platform integrates AlphaFold2-generated protein structures, MolGPT-driven scaffold generation, and a tailored ChatGPT model orchestrating Structure-Activity Relationship (SAR) analyses, virtual screening, and iterative optimization cycles. We applied this framework retrospectively to twenty FDA-approved small-molecule drugs spanning cardiovascular, neurological, oncology, and endocrine therapeutic areas. Each case study illustrates how AI systems can recapitulate historical discovery pathways with high fidelity, as demonstrated by metrics including structural similarity (average Tanimoto coefficient ≈ 0.82) and bioactivity-prediction concordance (mean Pearson r ≈ 0.78). The methodology emphasizes bioisosteric replacements, scaffold hopping, and pharmacophore optimization, reflecting human medicinal-chemistry strategies. The implementation of an AI-driven retrosynthetic platform, "ChemGPT Discover," exemplifies automation of medicinal-chemistry processes, enhancing efficiency in hit-to-lead development. Our results validate the capability of LLM-assisted retrosynthesis to rediscover known drug leads accurately, underscoring the transformative potential of AI in accelerating drug discovery and medicinal chemistry research.