Construction safety management has long struggled to produce daily safety bulletins (DSBs) efficiently. These bulletins must list site‑specific hazards, convey safety protocols, and propose preventive measures. Manual authoring typically consumes 2–3 hours per day, yields inconsistent quality, and about 25% of crews never perform pre‑task planning, pulling safety staff away from field supervision.
This paper introduces the Lightweight Multi‑Agent DSB System (LM‑DSB). The core combines retrieval‑augmented generation (RAG) with a locally deployable large language model (LLM) that runs on consumer‑grade PCs, achieving fully automated DSB creation. A hierarchical architecture coordinates three specialized agents—planner, researcher, and writer. The model is fine‑tuned on 37 health, safety, and environment (HSE) documents, and RAG grounds generation in retrieved passages from this knowledge base.
Validation proceeds in two phases. Phase 1 comparative tests show that the multi‑agent RAG configuration using Llama 3‑8B outperforms single‑agent and non‑RAG baselines, achieving TF‑IDF 0.70, BLEU 0.39, ROUGE‑1 0.66, and a document length ratio of 1.01 versus 0.39 for the single‑agent setup. Phase 2 involves 40 domain experts who rate overall acceptability between 3.50 and 4.20 on a 5‑point scale, with the highest scores in regulatory compliance (4.20) and factual accuracy (4.10).
The findings demonstrate that a well‑designed architecture enables lightweight models to handle safety‑critical documentation on consumer hardware, advancing democratization of construction safety technology. Limitations include the scope of the knowledge base, language constraints, and the absence of field deployment validation. Future work should explore multimodal integration and real‑world workflow adoption.
Review