Architected and directed AI-agent-driven development of CubeXplorer, an industrial AI reasoning system that reads a PLC program's control-logic export together with read-only runtime data, builds a knowledge-graph model of the machine, and diagnoses degradation and advises controller improvements — with hard architectural guarantees that it only ever reads from equipment and never acts autonomously.
Combines a knowledge graph, an out-of-process physics/estimator simulation, and a model-agnostic LLM interface. When a component isn't configured, the system degrades honestly rather than fabricating an answer — every claim it makes is traceable to reproducible evidence.