PermitGPT Unifies AI for Construction Governance and Impact Assessment

PermitGPT introduces a unified generative artificial intelligence framework designed to streamline urban construction governance. This system tackles the challenge of fragmented data by integrating information from various municipal and regulatory sources, such as the New York City Department of Buildings, OSHA, and NYC 311 service requests. By spatially and temporally aligning these records, PermitGPT creates a rich dataset of 90,000 structured prompt-response pairs. The core function of PermitGPT is to convert free-text construction permit descriptions into structured outputs across three critical domains: identifying potential safety hazards, specifying necessary permit requirements, and assessing community impact. The researchers fine-tuned three open-weight language models (Gemma-3-1B, Llama-3.2-3B, and 4-bit Mistral-7B-Instruct-v0.3) using parameter-efficient adaptation. Evaluation on a held-out test set demonstrated complementary strengths among the models: Gemma-3-1B offered the most efficient inference, Llama-3.2-3B excelled in lexical overlap for regulatory outputs, and Mistral-7B-Instruct-v0.3 achieved the strongest semantic alignment. While BLEU scores were low due to the open-ended nature of structured generation, semantic metrics and qualitative analysis confirmed the utility of the outputs. PermitGPT represents a significant initial step towards AI-assisted construction governance, highlighting future directions for robust task-level evaluation and real-world deployment.