Automated Compliance Checking (ACC) of building designs traditionally relies on formal- ized rules applied to structured Building Information Modeling (BIM) data. This thesis investigates whether an agent-based inspection workflow using rendered BIM snapshots and Vision Language Model (VLM) reasoning can support visual compliance checking as a bounded decision-support task. The proposed snapshot-grounded framework uses an inspection agent to iteratively collect rendered observations from a navigable BIM viewer, interpret them against a natural-language rule with a VLM, track unresolved evidence requirements, and finalize the assessment as compliant, non-compliant, or inconclusive. A browser-based prototype was implemented with evidence-driven navigation, VLM-based reasoning, trace persistence, secondary judge review, and report generation. It was evaluated on 27 controlled synthetic BIM configurations covering eight accessibility- and safety-related rule types derived from Solibri rule templates. In binary reasoning mode, Claude Sonnet 4.6 and GPT-5.4 achieved verdict accuracies of 40.7% and 37.0%, respec- tively, across 54 primary checks. Both models performed best in visually unambiguous presence-absence conditions but struggled with exact dimensional estimation, compound sub-checks, localized violations, and thin or closely bounded geometry. Further, trace analysis showed that some directionally correct verdicts were based on the incorrect sub-condition. Binary mode suppressed inconclusive outcomes that the available evidence would have otherwise justified. These results demonstrate the feasibility of snapshot- grounded, agent-based visual inspection as a traceable decision-support mechanism but reveal that it is not yet reliable for autonomous compliance checking. Therefore, the main contribution of this thesis is a transparent inspection framework that can system- atically diagnose failures in evidence collection, VLM reasoning, and secondary review through stored trace artifacts. The prototype implementation, evaluation test scenes, and inspection result reports are publicly available.     «

Automated Compliance Checking (ACC) of building designs traditionally relies on formal- ized rules applied to structured Building Information Modeling (BIM) data. This thesis investigates whether an agent-based inspection workflow using rendered BIM snapshots and Vision Language Model (VLM) reasoning can support visual compliance checking as a bounded decision-support task. The proposed snapshot-grounded framework uses an inspection agent to iteratively collect rendered observations from a navig...     »