"We have developed QEV, an AI decision-making model. Previously, systems detected sensitive information by using predefined code patterns for phone numbers, contact lists and resident registration numbers. QEV, however, uses its own training based on a large language model (LLM) to understand context and detect sensitive information and anomalies."
Kim Sam-young, chief technology officer (CTO) of QueryPie, made the remarks in an interview with MoneyToday on Nov. 10. He said QEV overcomes the limitations of conventional methods, which relied on predefined or previously learned categories to identify sensitive information, despite the use of various approaches ranging from regular expressions and keyword rules to statistical and machine learning classification.
The company described QEV as the first AI decision model of its kind developed in South Korea. Designed specifically for security in the data loss prevention (DLP) sector, the model aims to prevent companies from leaking sensitive information. QueryPie is seeking to gain a technological edge amid growing corporate demand for data protection.
According to Kim, QEV analyzes input text, classifies information and determines the optimal response. Unlike conventional generative AI, which focuses on answering questions or creating new content, QEV determines whether given information meets specific criteria and provides the most appropriate result from a set of predefined choices, along with the probability assigned to each option.
"QEV can identify whether a sentence contains personal information or confidential corporate information when an employee enters customer data or unreleased business materials into an AI service," Kim said. "It can detect potential information leaks before sensitive data is exposed externally and support reviews and responses in accordance with security policies."
QEV is built on Gemma 4, an open AI model, with QueryPie's proprietary decision-making technology added. The company further trained and optimized the model using corporate security data to develop a product specialized in detecting sensitive information.
"QEV can learn from company-specific data and work characteristics, so we expect it to be used in a wide range of environments," Kim said. "
The significance of this achievement is that we developed a model with state-of-the-art performance in the DLP sector by further training and optimizing an open AI model using QueryPie's proprietary technology and security-specific data."
In performance tests using internal corporate documents in Korean, English and Japanese, QEV achieved a sensitive information detection rate of 99.69 percent, identifying 1,277 of 1,281 documents containing sensitive information. The detection rate of JEV API, an externally available commercial decision-making model, was 96.41 percent, 3.28 percentage points lower than QEV's.
Kim also said QEV offers operational efficiency. Running AI models typically requires substantial computing resources and GPU memory, but QEV models reduced GPU memory usage by 27.5 to 45.2 percent while largely maintaining their sensitive information detection performance.
QueryPie plans to make QEV available on Hugging Face, an AI model-sharing platform.
It has also launched a separate demo page where users can easily try the model's decision-making capabilities. The demo supports Korean, English and Japanese, allowing users to check classification results and the probability assigned to each option using either preset test cases or their own text. By opening the service to the public, the company says it aims to let anyone verify the model's performance and potential applications.
"Building on the results of our DLP performance tests, we will continue to improve AI technology for practical corporate use and expand its applications beyond security into various enterprise sectors," Kim said. "The potential applications are virtually limitless."
Founded in Silicon Valley, the United States, QueryPie specializes in enterprise security and AI and aims to pursue an initial public offering (IPO) in South Korea.
In June 2020, the company officially launched its Access Control Platform (ACP), developed by engineers formerly with Kakao and Naver alongside experts from various fields. It has since expanded its business in access control and auditing across data, infrastructure and workplace systems.
More recently, the company has added AI capabilities to its security platform, helping businesses maintain consistent controls over permissions, approvals, auditing and sensitive information protection while using AI tools and agents.
in STOCK
