Source-linked AI summary
PermitGPT: A Unified Generative-AI Pipeline for Construction Hazard Forecasting, Permit Prediction, and Community Impact
Mohd Ruhul Ameen, Farjana Aktar, Akif Islam, Momen Khandoker Ope, Abu Saleh Musa Miah, Jungpil Shin
TL;DR
Construction-governance evidence is fragmented across safety, permitting, and community systems, complicating early assessment. PermitGPT aligns these sources and adapts three open-weight models to generate structured outputs across the three domains, with complementary trade-offs among efficiency, lexical overlap, and semantic alignment. The study presents an initial decision-support framework while identifying limits in evaluation and data annotation that require future validation.
Problem
Construction governance requires joint early assessment of safety hazards, permit requirements, and community impacts, but relevant evidence remains fragmented across separate systems.
Method
PermitGPT spatially and temporally aligns municipal and regulatory data, then uses parameter-efficient adaptation to fine-tune three open-weight models for structured decision support.
Results
The three models show complementary behavior: Gemma-3-1B favors efficiency, Llama-3.2-3B lexical overlap, and Mistral-7B-Instruct-v0.3 semantic alignment.
Takeaways & Limitations
PermitGPT provides an initial framework for AI-assisted construction governance in which different models may serve different operational needs.
Takeaways & Limitations
The study lacks zero-shot baselines, task-specific category accuracy, component ablations, and full double annotation, limiting attribution and formal agreement assessment.
Abstract
from arXiv · showhide
Urban construction governance requires early decisions that connect workplace safety, permitting requirements, and community impact, yet the relevant evidence is often scattered across separate municipal and regulatory data sources. This paper presents PermitGPT, a unified generative artificial intelligence framework for converting unstructured construction permit descriptions into structured decision-support outputs across three domains: safety hazard identification, permit requirement specification, and community impact assessment. To address data fragmentation, we spatially and temporally align records from the New York City Department of Buildings, Occupational Safety and Health Administration, and NYC 311 service requests, producing 90,000 structured prompt-response pairs derived through rule-based alignment and domain-informed spot checking. We fine-tune three open-weight language models using parameter-efficient adaptation and evaluate them on 2,833 held-out test cases. The results show complementary model behavior: Gemma-3-1B provides the most efficient inference at 3.07 samples per second with low memory usage, Llama-3.2-3B gives the highest lexical overlap for regulatory-style outputs with a BLEU score of 0.0091, and 4-bit Mistral-7B-Instruct-v0.3 achieves the strongest semantic alignment with a BERTScore-F1 of 0.7747. Because the task involves open-ended structured generation, low BLEU values are interpreted alongside semantic metrics and qualitative output structure rather than as standalone indicators of utility. Overall, PermitGPT provides an initial step toward AI-assisted construction governance while identifying directions for stronger task-level evaluation and real-world validation.
I. INTRODUCTION
PermitGPT addresses fragmented construction-governance evidence by converting permit descriptions into structured outputs spanning safety, permitting, and community impacts. It uses aligned data and adapted open-weight models as an early decision-support layer rather than a replacement for experts.
- Construction permit descriptions can signal worker hazards, regulatory requirements, disruptions, and community concerns, but these signals are rarely examined together.
- Dense urban construction makes fragmented safety and compliance evidence especially consequential for workers, residents, businesses, and infrastructure.
- PermitGPT converts raw construction permit descriptions into structured outputs for safety hazards, permit requirements, and community impacts.
- The pipeline spatially and temporally aligns heterogeneous municipal and regulatory sources before fine-tuning three open-weight language models.
- PermitGPT is designed as an early decision-support layer that organizes fragmented information without replacing human inspectors or regulatory experts.
II. RELATED WORK
Related work shows that generative AI can support construction safety and document-based decision support, but existing systems commonly target individual tasks. PermitGPT follows domain-adaptation approaches to compare open-weight models within a unified pipeline.
- Generative AI has supported construction-document interpretation, safety-record summarization, hazard detection, and compliance-oriented decision support.
- Prior systems include accident classification and multimodal inspection assistance, but generally address specific safety tasks rather than joint governance outputs.
- Domain-specific construction language motivates adapting general-purpose language models for permits, safety incidents, and regulatory documents.
- Parameter-efficient methods such as LoRA and quantization reduce adaptation costs while preserving useful task performance.
- PermitGPT uses parameter-efficient adaptation to compare three open-weight models under the same pipeline.
C. Evaluation, Ethics, and Responsible Use
The paper frames PermitGPT within the challenges of evaluating generative construction systems and deploying them responsibly. Its research gap concerns unified, fragmented-data assessment and practical lightweight models, while responsible use keeps human expertise central.
- Evaluation: Generative construction systems require evaluation beyond classification metrics because multiple valid textual responses can express the same risk.
- Ethics and Responsible Use: Responsible AI deployment must address regulations, privacy, fairness, accountability, and the limits of automated decision-making.
- Related Directions: Existing research offers opportunities through graph models, transfer learning, hybrid human–AI systems, and real-time monitoring.
- Research Gap: Three gaps are identified: single-task focus, fragmented early-stage evidence, and limited study of lightweight models for structured multi-domain outputs.
- Position of PermitGPT: PermitGPT provides an initial framework for connecting fragmented information while supporting future task-specific evaluation, baseline comparison, and real-world validation.
- Position of PermitGPT: The three model branches represent efficiency-oriented screening, regulatory-style lexical precision, and semantic depth and coherence.
III. METHODOLOGY
PermitGPT integrates permit, safety, and community-request data through spatial and temporal alignment to generate three categories of construction decision-support outputs. The corpus uses rule-based matching with domain-informed spot checking rather than full double annotation.
- The framework combines NYC Department of Buildings records, OSHA Enforcement Data, and NYC 311 service requests.
- The Buildings database supplies permit descriptions and structured project information, including project type, location, and contractor fields.
- OSHA records provide workplace inspection and violation context, while NYC 311 requests capture noise, obstruction, dust, and access-disruption signals.
- Rule-based spatial and temporal alignment links permits with nearby incidents and complaints during active permits and an additional 180-day window.
- Each aligned permit description is connected to likely safety hazards, possible permit requirements, and anticipated community impacts.
- Corpus construction used rule-based alignment and domain-informed spot checking rather than full manual double annotation.
B. Model Selection and Fine-Tuning
PermitGPT compares three open-weight models selected for different deployment trade-offs and adapts them with LoRA for construction-domain tasks.
- Gemma-3-1B represents lightweight deployment with rapid screening and lower memory usage.
- Llama-3.2-3B represents a mid-sized option targeting stronger lexical alignment for regulatory-style text.
- Mistral-7B-Instruct-v0.3 represents a larger instruction-tuned option targeting stronger semantic coherence and more detailed responses.
- LoRA updates a small set of trainable adapter parameters while keeping the original model weights frozen.The study sets the LoRA rank to r = 8 for construction-domain adaptation with limited computational overhead.
- Training uses AdamW, a 2 × 10−4 learning rate, 0.01 weight decay, warmup, and linear decay across the three models.The shared setup supports lightweight adaptation and comparability of performance and efficiency.
C. Output Structure and Evaluation Setup
PermitGPT converts permit descriptions into structured responses spanning safety, permitting, and community impact, then evaluates generation quality and deployment efficiency.
- Each model converts an input permit description into safety hazard, permit requirement, and community impact components.The structure is intended to organize early construction information for human review.
- 2,833 held-out test cases support evaluation of the generated responses.The evaluation considers both text-generation quality and inference efficiency.
- BLEU, ROUGE-1, and METEOR measure lexical or surface-level similarity, while BERTScore measures semantic alignment.Inference efficiency is included to assess practical deployment trade-offs among the models.
A. Quantitative Performance Evaluation
The evaluation reports complementary model behavior: Llama-3.2-3B leads lexical metrics, Mistral-7B-Instruct-v0.3 leads semantic alignment, and BLEU requires cautious interpretation for open-ended generation.
- Llama-3.2-3B achieves the highest test BLEU, ROUGE-1, and METEOR scores.This indicates relatively stronger surface-level overlap with reference responses.
- Mistral-7B-Instruct-v0.3 achieves the highest test BERTScore, indicating stronger semantic alignment despite differing wording.
- BLEU and METEOR are less informative here because multiple valid structured responses can express the same meaning with different wording.BERTScore is treated as the main semantic-alignment indicator because contextual embeddings tolerate valid paraphrase variation.
- The metrics remain reasonably stable across training, validation, and test splits without large split-level fluctuations.The study compares three fine-tuned open-weight models but does not include a separate zero-shot baseline.
B. Qualitative Output Analysis
Qualitative analysis frames PermitGPT as a structured early decision-support tool, while deployment results highlight a trade-off between efficiency and semantic alignment.
- Output Structure: Table II illustrates the organization of generated outputs into safety, permitting, and community-impact categories.The example demonstrates practical output organization rather than serving as an additional quantitative benchmark.
- Output Structure: The structured response can help reviewers identify issues requiring closer human inspection.
- Deployment Considerations: Gemma-3-1B reaches approximately 3.07 samples per second with comparatively low VRAM usage.This makes the lightweight model attractive for rapid screening or preliminary review scenarios.
- Deployment Considerations: Mistral-7B-Instruct-v0.3 provides stronger semantic alignment but requires higher computational resources.The comparison reflects a trade-off between model capacity and inference cost.
- Deployment Considerations: The models support different operational needs because deployment depends on balancing speed, memory use, consistency, and output quality.Llama-3.2-3B occupies a middle position with stronger lexical overlap for regulatory-style responses.
- Deployment Considerations: Fine-tuning loss decreases smoothly without major instability, supporting the adequacy of the LoRA adaptation strategy in this setting.
D. Human-in-the-Loop Use and Quality Assurance
PermitGPT is intended to support human inspectors and regulatory experts by organizing permit information into reviewable hazard, permitting, and community-impact outputs. The study also identifies evaluation and data-construction boundaries that limit how its results should be interpreted.
- Human-in-the-Loop Use: PermitGPT supports rather than replaces human decision-makers in construction safety and permitting workflows.Generated outputs are intended as decision-support information for review and refinement by inspectors or municipal staff.
- Quality Assurance: Confidence-aware routing and manual verification can improve accountability for incomplete, overly general, or uncertain model responses.Lightweight models may screen routine projects, while complex cases can receive detailed model outputs or human review.
- Evaluation Boundaries: The study compares three fine-tuned open-weight models without zero-shot baselines, task-specific accuracy, or component-level ablations.Consequently, the separate effects of fine-tuning and individual data sources cannot be quantified from the reported results.
- Data Quality: The corpus used rule-based alignment and domain-informed spot checking rather than full double annotation, so formal inter-annotator agreement was not measured.This limits direct assessment of annotation consistency.
- Future Directions: Future work will add controlled baselines, structured task-level evaluation, component-wise ablations, and prospective real-world case studies.The paper also proposes multimodal inputs, real-time sensor data, and federated learning for broader deployment evaluation.
V. CONCLUSION
PermitGPT organizes fragmented construction information into structured decision-support outputs spanning safety, permitting, and community impact. Its three-model evaluation shows complementary behavior across efficiency, lexical overlap, and semantic alignment, while positioning the system as an early-stage assistive tool requiring further validation.
- V. CONCLUSION: PermitGPT integrates permit records, safety data, and community service requests across three construction-governance concerns.These concerns are safety hazard identification, permit requirement specification, and community impact assessment.
- V. CONCLUSION: Three fine-tuned open-weight language models show complementary behavior across efficiency, lexical overlap, and semantic alignment.The paper suggests that different models may serve different operational needs.
- V. CONCLUSION: PermitGPT is intended as an early-stage assistive tool that organizes information for review rather than replacing human inspectors or regulatory experts.The current study is an initial evaluation of this direction.
- V. CONCLUSION: Further work includes zero-shot baseline comparison, task-specific accuracy measurement, component-wise ablation, and real-world case validation.These gaps remain part of the paper’s stated future research agenda.