Towards Full-stack AI for Chemical Synthesis Planning

This talk presents a full-stack AI Agent for chemical synthesis planning, connecting the construction of chemical metric space, single-step retrosynthesis, strategic multi-step route generation, and data-efficient experimental validation. The goal is to make synthesis planning not only more accurate, but also more practical and reliable for real-world chemistry. The framework combines chemically informed representations, structure-aware retrosynthesis models, LLM-assisted multi-step planning, and feasibility assessment based on wet-lab data. By integrating these layers, the system can propose routes, reason over alternatives, reduce invalid or unrealistic steps, and prioritize experimentally robust solutions. Overall, the talk argues that the future of synthesis AI lies in combining strong chemical inductive bias with scalable AI planning and real experimental feedback, moving toward industry-deployable synthesis intelligence. Parts of this project have been deployed at leading CRO and AI4Chem company.