A Human-Supervised Multi-Agent Decision-Support System for Safety Non-Compliance Reporting from Construction-Site Images

Creative Commons License

Semercioğlu Akbulut İ. N., Başağa H. B., Hacıefendioğlu K.

9th International Project and Construction Management Conference (IPCMC2026), Gaziantep, Türkiye, 16 - 18 Ekim 2026, ss.171-182, (Tam Metin Bildiri)

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Basıldığı Şehir: Gaziantep
  • Basıldığı Ülke: Türkiye
  • Sayfa Sayıları: ss.171-182
  • Karadeniz Teknik Üniversitesi Adresli: Evet

Özet

This study develops and initially evaluates a human-supervised multi-agent artificial intelligence decision-support system to assist occupational safety specialists in identifying and reporting safety non-compliances from construction-site images. The proposed workflow integrates visual data acquisition and characterization, visual scene and hazard interpretation, interim expert review, regulatory mapping, corrective and preventive action (CAPA) recommendation, structured report generation, and final expert approval. The system was implemented entirely in a local computing environment, with the Qwen2.5-VL:3B vision-language model executed through Ollama for visual interpretation and a keyword-based local search used to retrieve relevant regulatory excerpts. To examine both the behavior of the agent chain and the contribution of interim professional review, the same construction-site image was processed under two conditions while the upstream agent outputs were kept identical. Without interim expert revision, the visual assessment prioritized secondary site elements, resulting in more heterogeneous regulatory retrieval and general CAPA recommendations. Following expert revision, the open floor opening and associated fall hazard were prioritized, the retrieved regulatory excerpts became more focused on fall- and edge-protection provisions, and the CAPA recommendations addressed the principal site condition more directly. The findings demonstrate the operational feasibility of the integrated multi-agent workflow and illustrate the value of retaining professional oversight within AI-supported construction-safety decision making. As the comparison was conducted using a single construction-site image, the findings represent an initial illustrative evaluation rather than a quantitative validation of system performance.