AISA - AI Safety Assistant Framework for Continuous Improvement of Highway Construction

Sep 15, 2026

Uploaded by Ivy

Job Safety Analysis (JSA) and pre-task planning improve highway construction safety, but historical accident data often exists as unstructured narratives that planners cannot readily consult during daily preparation. This work proposes an AI Safety Assistant (AISA) framework using large language models with deterministic local inferencing. The approach supports incident narrative classification and quality scoring, and retrieves relevant past accidents, imagery, and trusted documents to strengthen safety plans through affordable, auditable, privacy-preserving local tooling.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages17

Reading Time43 min

AccessFree to download

Abstract

1 Introduction

Job Safety Analysis and the need for historical records

Progress in safety analytics and NLP-based classification

Safety assistants and agentic approaches

AISA: AI SAFETY ASSISTANT FRAMEWORK FOR CONTINUOUS IMPROVEMENT OF HIGHWAY CONSTRUCTION arXiv : 2608 . 17 184v 1 [cs .CL] 17 Aug 2026 Mason Smetana Department of Civil and Environmental Engineering University of Pittsburgh Pittsburgh, PA 15213 [mrs196@pitt.edu](mrs196@pitt.edu) Trevor Neece Department of Civil and Environmental Engineering University of Pittsburgh Pittsburgh, PA 15213 [trevor.neece@pitt.edu](trevor.neece@pitt.edu) Lev Khazanovich Department of Civil and Environmental Engineering University of Pittsburgh Pittsburgh, PA 15213 [lev.k@pitt.edu](lev.k@pitt.edu) August 1, 2026 ABSTRACT Objectives: Job Safety Analysis (JSA) and pre-task planning can benefit from prior incident records, yet historical accident data is often stored as unstructured narratives that are difficult to consult atthe point of planning. A novel framework centered on large language models (LLMs) for highway construction safety reporting and planning is proposed as a foundation for future agentic applications, prioritizing deterministic, local inferencing. The first aim is to enable classification and quality scoring of incident narratives for existing and future reporting purposes. The second is to evaluate retrieval of relevant historical accidents, related imagery, and trusted industry documents for incorporation into daily safety plans. Methods: Neural probes were trained to classify incidents along four multiclass and two binary Occupational Injury and Illness Classification System (OIICS) fields and to derive an overall quality score, evaluated on a test set of over 15,000 narratives and a held-out set of 100 author-labeled records, benchmarked against a majority-vote LLM ensemble. The retrieval of historical accidents, reference imagery, and industry documents was benchmarked across embedding models using standard information retrieval metrics. Findings: OIICS classification reached 75% held-out accuracy, though the two binary flags were degenerate. The quality score, while meaningful on one database, was distorted on out-of-distribution fatalities in the held-out dataset. Accident retrieval recovered relevant incidents far above chance, performing best on lexically distinct construction activities. On document question answering, an open-weight decoder embedding model surpassed proprietary models. Novelty: Overall, this work provides a new framework rooted in local inferencing and text embedding models for future agentic applications, with emphasis on bridging external data to JSA reports. Practical Applications: The findings guide practitioners toward affordable, auditable, privacypreserving safety tooling that runs locally and folds new incidents and near-misses back into planning, without transmitting sensitive records to cloud providers. 1 Introduction Highway and heavy civil construction work concentrates some of the most severe hazards in the American workplace [1] . In just a few examples, workers operate near heavy equipment, erect and dismantle formwork, and perform maintenance in close proximity to live traffic. The primary defense against these hazards is not any technology alone, but a practice of Job Safety Analysis (JSA) and pre-task planning, in which a safety professional breaks an upcoming task into steps, identifies hazards at each step, and selects controls before work begins [2] . The quality of said analysis depends almost entirely on whether the planner foresees potential hazards for a given work activity. Historical incident records, on one hand, provide the necessary foresight for dangers presented in similar work activities [3]; yet such repositories may be difficult to use or consider during the planning phase. They exist as thousands of unstructured accident narratives (free-text descriptions embedded within otherwise structured records), indexed by report metadata (e.g., dates, establishments, injury codes) . Thus, connecting JSA to historical data remains a largely manual an

FAQs

Why are historical accident narratives hard to use in job safety analysis?

They are stored as thousands of unstructured free-text narratives that are difficult to consult during the planning phase, even though they contain the foresight needed for similar work activities.

What are the two main objectives of the AISA framework?

First, classify and score incident narratives for reporting purposes. Second, retrieve relevant historical accidents, related imagery, and trusted industry documents to incorporate into daily safety plans.

How does AISA evaluate its methods and performance?

It trains neural probes for OIICS incident classification and quality scoring, evaluates on a test set of 15,000+ narratives and a held-out set with author-labeled records, and benchmarks retrieval using standard information retrieval metrics.

Sep 15, 2026

Uploaded by Ivy

AISA - AI Safety Assistant Framework for Continuous Improvement of Highway Construction_Page 1

Open in WPS to view all 17 pages