Built on a four-layer architecture spanning Interaction, AI, Acquisition, and Execution: the Interaction layer provides a digital-twin P&ID alongside dual voice/text command channels; the AI layer leverages multi-agent orchestration and Bayesian optimization to propose subsequent experimental conditions; and the Acquisition layer aggregates temperature, pressure, and flow data via industrial PLC and DAQ into a time-series database. Operators can issue valve, MFC, and temperature-control commands directly in natural language—validated against safety thresholds prior to execution and overseen by multi-level safety interlocks if limits are exceeded. Online GC feeds back conversion and selectivity in real time to close the optimization loop, enabling 24/7 unattended catalyst screening and process optimization.