Generative AI co-pilot for dynamic bowtie diagrams in construction risk management

Abstract

This thesis presents the design, implementation, and evaluation of Nora, an AI generative co-pilot developed to support the construction and refinement of bowtie diagrams in the field of industrial risk management. Bowtie diagrams are structured graphical models that illustrate how specific threats can lead to an undesired top event and, subsequently, to potential consequences, while also showing the preventive and mitigative barriers in place to control both threats and consequences. The work was conducted at Beawre Digital S.L., a technology start-up dedicated to digital risk automation and continuous risk management for the construction industry. Traditional bowtie tools are effective for visualization but require extensive manual effort, limiting knowledge reuse and scalability. Nora addresses these limitations by combining large language models (LLMs), embeddings, and retrieval-augmented generation (RAG) into a single interactive assistant that communicates through natural language and updates diagrams dynamically. The system was re-architected from a slow, error-prone prototype into a scalable, production-ready service. With this new approach, task or intent classification and context retrieval occur simultaneously, followed by real-time response streaming and persistent "smart view" storage in Azure Blob Storage. A multi-agent orchestration layer separates classification, generation, and validation tasks, ensuring linguistic consistency and robust recovery from errors. In parallel, a weighted embedding framework allows semantic comparison of bowtie diagrams, supporting knowledge reuse across projects. The entire pipeline is deployed securely through GitLab CI/CD, Docker, and Kubernetes for automated scalability. The empirical evaluation compared the improved and initial versions of Nora across three dimensions: correctness, latency, and cost. The results show significant improvements: mean latency decreased fivefold; correctness improved from 59% to over 93%; and cost per request was reduced by up to 70%. Statistical testing confirmed that these improvements were both reliable and significant. These advances stem from concurrent processing, structured and extensive prompt engineering, and asynchronous embedding computations. Conclusively, the thesis demonstrates that generative AI can operate reliably within structured industrial frameworks when supported by modular orchestration, validation, and semantic grounding. Nora thus represents a novel blend of conversational AI and formal safety methodologies that supports Beawre's vision of continuous, data-driven risk management. Future research will focus on multi-agent coordination, enhanced data governance, and the transition toward hosting and fine-tuning an in-house language model to ensure long-term independence and scalability.