Hee-jeong Bae (배희정), CEO of KMS Lab.

South Korean software company KMS Lab is working on POSCO’s next-generation enterprise portal project, Next EP, based on its self-developed AI knowledge platform, K-Portal BrainWeave.

KMS Lab said the project is highly significant for the company and for the broader domestic industry because the small business won the contract after running a proof-of-concept race for more than 7 months against well-known global and domestic AI companies. The project replaces an internal portal used by tens of thousands of people with an AI-based system.

KMS Lab CEO Hee-jeong Bae (배희정) also stressed at a recent press briefing that many companies sell standalone AI products such as meeting-minute writing or email summarisation. She said she had not heard of a case in which a small company deployed a full-stack AI solution on a large group-wide system. She again emphasised that the company supplied the entire software stack with its own technology, from model serving and training data processing to business applications.

KMS Lab’s AI knowledge platform, K-Portal BrainWeave, deployed at POSCO is an AI-based knowledge management solution that turns individual work experience and expertise into a shared organisational asset. It supports continuity and productivity by systematically accumulating and sharing knowledge scattered within a company.

POSCO’s Next EP is an on-premise project that runs large language models (LLMs) on in-house GPUs without sending data to an external cloud. It connects and uses multiple open-weight models specialised for specific purposes.

This may be a case showing that while large companies also use ChatGPT or Gemini, they prefer to handle emails and key materials that must not leak inside their own systems.

But it is not as easy as expected for on-premise LLMs to settle into enterprise workplaces. Performance is cited in particular as a stumbling block. With business users already calibrated to cloud-based global frontier AI, on-premise LLMs with weaker performance do not easily gain traction on the ground simply because they are safer.

KMS Lab also focused on addressing such issues as it took on the POSCO project. Bae explained that the company was able to secure POSCO as a client because it properly addressed the pain points of on-premise LLMs.

She said, "Even if it cannot be as good as OpenAI or Claude, employees will not use it unless it comes close." She added that it received a good evaluation for providing open-weight LLMs on-premise with performance close to cloud LLMs.

It may sound easy, but building an environment in which on-premise LLMs can be used at frontier AI levels requires substantial capability. It also does not work simply by bringing in open-source LLMs with good performance.

In enterprises, LLMs go beyond the models themselves and include connections to internal corporate data and existing systems. Success or failure depends not on the LLM alone but on how well it chemically combines with existing systems, Bae said. She also said that how LLMs are integrated and connected has a big impact on performance in ways that are not always obvious.

In this regard, Bae emphasised KMS Lab’s data processing capabilities. She said the company’s accumulated technologies related to data labelling, cleansing, classification and deduplication are serving as a competitive advantage that can be felt when deploying on-premise LLMs in enterprise workplaces.

Bae said, "We did not anticipate and prepare for generative AI, but know-how built up over 26 years worked as a strength in cleansing training data." She said this helped raise the accuracy of LLM responses. She also said Meta effectively acquired Scale AI for a large sum because it highly valued data processing capabilities.

One of the core products behind the data processing capabilities Bae repeatedly emphasised is an archiving solution that stores documents and emails in real time. Because it even retains emails users delete, the volume of data accumulated through archiving is enormous. If it is stored as-is, networks and storage struggle to handle it. KMS Lab has therefore provided an archiving solution focused on deduplication, compression and classification.

Bae said, "In a 2016 external institution test, we reduced 1 terabyte that included images by 73.5 percent, and now we have raised it to around 80 percent."

KMS Lab also devoted effort to integration with existing systems as it carried out the POSCO project. Bae said, "Because this is work to integrate and optimise systems and data, one company cannot do 10 or 100 at the same time. Even so, experience from doing it first creates a gap." She said it could also provide its self-developed software stack through a cloud approach to small and mid-sized companies that cannot afford to buy GPUs. She added, "We earned money without investment and continued research and development." She said the company improved by learning what it lacked on the ground and is still learning. She added that next year it will expand into agentic AI and physical AI being researched at the University of Toronto, along with an overseas expansion.

Keyword

#KMS Lab #POSCO #BrainWeave #Next EP #LLM
Copyright © DigitalToday. All rights reserved. Unauthorized reproduction and redistribution are prohibited.