An architectural drawing set looks like a stack of images, but its real content is a web of relationships: a concrete spec on sheet A-187 governs a footing on sheet S-4, and one wall note implies three finished surfaces downstream. Process the set as raster and every link is lost. This is the same failure spatial-reasoning research keeps measuring, where baseline LLMs score around 0.37 F1 on simple spatial questions and fall to roughly 0.08 once a query spans multiple concepts, because embeddings flatten geometry into tokens.
In this session, Michael will show you how he extracts a drawing set into Neo4j as what it physically is: a building's anatomy of structure, MEP, envelope, and the specifications that govern each. You will learn how to structure an ontology a language model can traverse without inventing edges, using first-class nodes that carry their own geometry, typed topological relationships, and a separation that lets the model propose while deterministic code disposes. You will also see how a bitemporal layer records both when something is true on site and when the graph learned it, and how "ghost" nodes stay provisional until field evidence confirms them.
Expect live Cypher that walks spec-to-geometry paths a vector index would miss, plus an honest account of the entity-resolution problems that were hardest to solve. You will leave able to model a spatial domain as a graph that reasons rather than retrieves.
Speaker
Mike Sturgeon
Founder & Systems Architect, VIA
Michael Sturgeon is the founder of VIA, a construction-verification platform built on a Neo4j knowledge graph. He spent roughly twenty years in construction field operations and stone fabrication before teaching himself to build software, driven by a problem he kept hitting: architectural drawings are the legal record of a building, yet computers read them as flat images and lose every relationship that matters. He now designs spatial knowledge graphs and agentic extraction pipelines that let language models reason over construction documents, now being validated on live jobsite scans in Denver.
