Permit parsing automation: 80% faster processing

An EdTech client approached Nexority Infotech with a critical need to automate their permit parsing process. Their existing manual system was slow, error-prone and unable to handle the complexity of nested hierarchies and unstructured data.

• Handling nested hierarchies: extracting data from deeply nested sections while maintaining context.
• Improving accuracy: high accuracy in data extraction despite unstructured and low-quality documents.
• Balancing performance and scalability: processing large volumes of permits quickly without compromising accuracy.
• Integration with existing systems: structured outputs compatible with the client’s data pipelines.

Dataset preparation

Permit PDFs were converted into images and annotated with LabelMe to create a diverse training dataset of sections, subsections and tables.

Model selection and fine-tuning

The Faster R-CNN model from Detectron2 was chosen and fine-tuned on the dataset, adjusting hyperparameters such as learning rate and batch size for improved accuracy.

Custom parsing logic

Regex-based logic extracts data from nested hierarchies and complex document structures.

Integration with the permit parser

The trained model was integrated into the NexoParser app to automate data extraction, outputting structured results in JSON and Excel formats.

• Time savings: manual processing time was reduced by 80%, letting the team focus on high-value work.
• Increased accuracy: data extraction accuracy improved to 93%, minimising errors in nested and complex information.
• Faster processing: documents are processed four times faster, so the client handles more permits in less time.
• Cost savings: automation reduced operational costs by 30%.
• Enhanced usability: structured JSON and Excel outputs simplified analysis and reporting.