LG AI Research: AEGIS

Quality control at manufacturing sites is directly linked to productivity, but AI model generalization has faced limitations due to changing process environments and complex defect criteria that rely on expert intuition. Existing AI models often suffered from sharp performance drops depending on the site, or created a cognitive gap where experts could not understand the basis for the AI's judgments.

To address this, LG AI Research introduced the AEGIS platform. AEGIS focuses on integrating domain knowledge that data alone cannot capture, building an environment where AI and on-site experts can collaborate.

The core elements of AEGIS are as follows:

• Providing explanations for anomalies: Rather than simply presenting probability values, it uses LLM-based agents to analyze historical records and domain documents, explaining the basis for judgments in natural language.
• Incorporating expert feedback: Experts verify the AI's judgments, and the domain knowledge gained in that process is fed back into the model's training, continuously evolving the system.
• Integration of domain knowledge: The system is designed to understand the experience and intuition of on-site experts that are not explicitly captured in the data.