Privacy-Preserving AI for Document Understanding with Controlled Unclassified Information

In manufacturing, manual exchange and review of engineering drawings constrains the supply chain. AI techniques have the potential to streamline these processes. This paper presents two important use cases and proposes solutions powered by generative AI models, including text and multimodal Large Language Models. The solutions preserve the privacy of documents used for training and inference and are thus suitable for controlled data. Performance is measured and discussed in the context of the applications, including technical and other practical considerations. Results indicate generative AI is comparable or better than previous solutions.