Every plant engineer has lost hours tracing dense P&ID drawings to answer one simple question: "What connects to this pump, and which valves isolate it?" These diagrams encode the connectivity of an entire facility — yet as static images, that knowledge stays locked away.
In this session, Hiroki will show how his team at JGC turned P&IDs into a living knowledge graph in Neo4j, and built an agent — powered by the Claude Agent SDK and a read-only Neo4j MCP server — that takes a plain-language question and writes its own Cypher. Instead of replying with text in a chat, the agent renders its answer interactively in the application's own UI, highlighting the exact equipment and connections on the diagram itself — no Cypher required from the user.
You will learn how a schema-based system prompt — not fine-tuning — grounds Cypher generation to avoid hallucinations, and how the agent returns stable node identities rather than pixel coordinates, so every highlight doubles as a citation for its answer. He will close with a live demo: ask "what gets isolated if this valve fails?" and watch the affected valves and lines light up across several linked drawings in seconds.
Whether or not you have ever seen a P&ID, you will leave with a reusable pattern: natural language, an agent, a graph, and a visual, explainable answer rendered on the diagram you already work with. The approach generalizes to any connected system you can draw — networks, supply chains, or circuits.
Speaker
Hiroki Kokubo
Digital Engineer, JGC Corporation
Hiroki Kokubo is a Digital Engineer at JGC Corporation, a global engineering, procurement, and construction company. For over three years he has led the development of Tag Graph Viewer, a Neo4j-powered system that turns complex plant documentation — especially P&ID drawings — into an interactive, queryable knowledge graph. His recent work explores natural-language search over engineering data, rendering the results directly on the diagrams engineers use every day. He is passionate about making hard-won engineering knowledge accessible through graphs. This will be his first NODES talk.
