NEXUS: A Multi-Agent Agentic Workflow for Valid SysML v2 Model Generation from Unstructured Industrial Documentation

The increasing complexity of modern cyber-physical systems necessitates Model-Based Systems Engineering, yet the practical adoption of SysML v2 remains hindered by high manual effort and steep linguistic learning curves. Existing AI-assisted approaches address only clean natural-language prompts or fragmented model fragments rather than the unstructured multi- page technical PDFs prevalent in industrial practice. This paper presents NEXUS (Neural EXtraction for Unified Systems engineering), a multi-agent agentic workflow for the semi-automated generation of formally valid SysML v2 models directly from unstructured technical documentation. The architecture decomposes the modeling process into three specialized agents: an Extraction Agent that transforms legacy PDFs into structured JSON representations of system hierarchy, connectivity, and behavior; a Building Agent that incrementally constructs SysML v2 code across three distinct phases with compiler-in-the-loop validation; and a Verification Agent that ensures semantic and structural integrity. Evaluation on a corpus of eight industrial-grade documents demonstrates that NEXUS achieves 100 % syntactic validity across all configurations, compared to 38 % for a state-of-the-art one-shot baseline, while reducing hallucination by a factor of 2.4. The proposed approach generalizes robustly across open-weight and proprietary backbone models, and incorporates human-in-the-loop oversight at every pipeline stage, providing a scalable and verifiable pathway for industrial MBSE adoption.