Research

I started in April 2026. This page describes the problem and what I am working on now. There are no results yet.

Problem

Performance-based design sits alongside prescriptive codes rather than replacing them. Most projects follow prescriptive guidance and turn to engineering analysis where that guidance does not fit the building, or to justify a specific departure from it. The two are usually combined in a single project.

Either route ends in an argument submitted to a regulator, reviewed by other engineers and carried under professional liability. That changes what an acceptable model output is. A fluent, confident, unsupported answer is worse than a refusal: it can reach a fire strategy, and fluency reads as competence to anyone without the source to hand.

Current work

• Measuring the reliability of language model output on fire-safety guidance: faithfulness to retrieved sources, claims made without support, and whether a cited clause says what the answer attributes to it.
• Retrieval over documents whose structure carries meaning — clauses, tables, cross-references — and how a system should behave when what it retrieves is not enough to answer.
• A risk-informed taxonomy of autonomy: which tasks in a workflow can be acted on directly, which need checking, and which should not be automated at all.

Domain and tools

The group's simulation tools are buildingEXODUS, for evacuation, and SMARTFIRE, for fire and smoke CFD. Both take substantial manual setup and produce output that has to be interpreted rather than read off. On the building-description side I work with openBIM/IFC and IfcOpenShell, and with graph representations — Neo4j and RDF.