Custom AI Agent Boosts Transportation Engineering with LLMs

New research details a structured methodology for creating specialized generative AI agents tailored for the transportation engineering sector. While general-purpose large language models (LLMs) show promise, their effectiveness in highly technical fields like transportation engineering is often limited by a lack of exposure to specific standards, terminology, and domain semantics. This study addresses this by proposing a systematic development framework. The approach involves curating a comprehensive corpus of U.S. transportation manuals, design guidelines, and regulatory documents. This specialized data is then used for continued pre-training of several state-of-the-art LLMs, employing a unified low-rank adaptation (LoRA) framework. The evaluation, using standard NLP metrics, showed that models such as Qwen2.5-7B and LLaMA-3.1-8B achieved superior domain alignment and response quality after this adaptation. This work provides a reproducible guideline for building domain-specific generative AI agents, paving the way for their broader application in transportation research, design, planning, and policy analysis.