An AI Assistant for Transportation Construction Standards and Decision Support

Musfira Mohamed

College: College of Business and Public Management

Major: BS.MANAGEMNT-SUPPLYCHAINMGT

Faculty Research Mentor: Liu, Dan

Abstract:

Urban vehicle routing and parking systems must follow complex and frequently changing regulations, which are often written in unstructured formats such as PDFs or semi-structured policy documents. When these rules are not interpreted or enforced correctly, routing systems may suggest actions that are illegal or impractical. This research focuses on developing and implementing a Policy-Aware Feasibility Pipeline that combines large language models (LLMs) with geospatial transportation data to ensure that vehicle routing and parking decisions always comply with real-world policy and infrastructure constraints. The proposed system uses an LLM-based policy parser to extract rules, conditions, and exceptions from unstructured urban regulatory documents and organizes them into a queryable knowledge graph. These structured policies are then integrated with GIS-based street and curb network data to create time-indexed feasibility matrices that represent legal and physical constraints. A key contribution of this work is incorporating these constraints directly into learning-based routing models through element-wise logit masking, which prevents the system from considering actions that are not legally or physically allowed. By enforcing policy compliance within the decision-making process rather than filtering results afterward, this research aims to improve routing reliability, regulatory compliance, and computational efficiency while maintaining high accuracy and low-latency performance.