[E-Daily Reporter Shin Yeong-bin ] Data and know-how accumulated in factories are being combined with generative AI to drive maintenance and research and development. LGELECTRONICS(066570)has introduced “Chada,” an Agent AI that supports R&D using in-house data, while POSCO has implemented a customized language model trained on equipment know-how.

The two companies presented case studies on the use of AI in manufacturing at the “Microsoft Industry Summit” held at COEX in Seoul on the 30th. They introduced examples of how they connected internal data and on-site knowledge to AI and applied it to actual operations, such as research and development and equipment management.

Adding Knowledge Exploration to Data Analysis… “Chatda” Performs Tasks

Lee Hyun-gu, a manager at LGELECTRONICS, introduced the development process of “Chatda,” the company’s internal data and knowledge agent. Chatda supports research and development tasks by linking internal data such as product usage history and design, testing, and quality documents.

Before the introduction of ‘Chada,’ analyzing data required requesting materials from the relevant department, verifying table structures, and then processing the data using SQL and Python. Lee said, “Previously, it took an average of about two weeks to obtain data analysis results, but with ‘Chada,’ we were able to reduce this process to approximately 45 seconds.”

“Chada” evolved from Version 1.0, which focused on structured data analysis, to Version 2.0, which integrated unstructured knowledge such as design, testing, and manufacturing records. In Version 3.0, data analysis and knowledge exploration were integrated into a single conversational flow. When a user asks a question, the AI formulates a plan, selects the necessary data and tools, and performs the analysis. Repetitive tasks are saved as “skills” so that the team can reuse them.

Lee explained, “Initially, we had 38 monthly active users, but that number has recently grown to 847,” adding, “This growth was driven not only by model performance but also by the expanded scope of data and tasks accessible to users.”

GitHub Copilot was also applied to the development process of “Chada.” The Copilot agent was designed to participate not only in writing code but also in design, task planning, and managing development history. Lee revealed that during the development of Chada 3.0, the team saved approximately 200 million won in outsourcing costs and shortened the development period by about two weeks.

POSCO, Facing Rising Retirements, Embeds Equipment Know-How into AI

Jang Won-jong, a team leader at POSCO, presented a case study on building a customized language model (LLM) for equipment using Microsoft’s AI Foundry. The impetus behind POSCO’s push for equipment AI stems from the challenge of passing on the experience and knowledge of maintenance personnel to the field.

“We have hundreds of retirees every year,” said Team Leader Jang. “How to formalize the know-how of on-site maintenance personnel is a challenge facing many companies.” He projected that the proportion of veteran maintenance personnel would drop from about 50% in the 2020s to below 30% in the 2030s.

POSCO applied fine-tuning to improve the response quality and utilization rate of its existing generative AI for equipment management. The company created training data from “POSWiki”—which contains POSCO-specific acronyms and equipment knowledge—and internal documents, while some question-and-answer data was generated by the AI.

The fine-tuned model was integrated with an internal document search function. Using Azure Document Intelligence, it analyzes materials such as on-site photos and graphs, and displays reference documents and sources alongside the answers. A system was also established to generate responses based on approximately 170,000 pieces of equipment-related content and to incorporate new documents into the training as they are registered.

The system is reportedly being used by new engineers to look up senior engineers’ maintenance experiences and checklists before inspecting equipment they are encountering for the first time. Currently, an average of about 100 inquiries are received daily. Team Leader Jang said, “The initial on-the-job evaluation score was around 75 points, but we improved the functionality and raised it to around 90 points.”

POSCO plans to expand its application of this technology from the equipment sector to the energy and materials sectors. Team Leader Jang said, “We are developing it into an integrated agent that links content with actual equipment status data to provide more precise guidance on equipment condition and appropriate responses.”