Developer Uses LLM to Turn Messy Supplier PDF Catalogs Into Airtable Data

A small e-commerce operator found that manually transcribing SKUs, descriptions, and prices from inconsistent vendor PDF catalogs was slow and error-prone. Traditional Python extraction libraries and template-based tools worked only for a minority of catalogs and required constant upkeep. The developer instead extracted raw text with pypdf and fed it to a self-hosted Llama 3 70B model, prompted to return structured product data. The approach reportedly reached about 90 percent accuracy, with human review covering the rest, and the resulting JSON was pushed into Airtable via a Python script.