Year: 2026
Type: article
Abstract: Home inspection evaluates the visible components of a residential building and reports their deficiencies, safety concerns, and recommended actions. Vision–language models (VLMs) can describe what a photograph shows, but inspection findings also require regulatory knowledge, system-level assessment, and familiarity with inspection practice. We present MASHI, a multi-agent vision–language system that answers home inspection questions across the ten systems of the ASHI Standard of Practice. A planner agent routes each question to an ASHI system and assigns tasks to a perception agent and a system agent. The perception agent reports visual evidence from the question’s photograph. The system agent applies the agent skill of the selected ASHI system to the question, its task, and this evidence. When needed, it retrieves text and images through multimodal retrieval-augmented generation (RAG) from a knowledge base of 45 building code editions and 18 inspection reports. For evaluation, a semi-automated pipeline with human verification turns 84 other inspection reports into 32,554 question–answer pairs. Prometheus-2 and Phi-4 rate the answers on a five-level rubric. MASHI reaches a mean judge score of 4.083, higher than those of all other compared systems, including its single-agent variant. Every difference is statistically significant.
DOI: 10.1145/3859927
Institutions: University of Virginia, Carnegie Mellon University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Funder: National Science Foundation
OpenAlex ID: W7225459443