QueryPie Develops QEV Decision-Making AI Model, Achieving 99.69% Sensitive Data Detection

A high-performance AI model built with proprietary AI technology and training and optimization using security-specific data

QueryPie, an enterprise data security and AI platform company led by CEO Brant Hwang, announced that it has developed its own decision-making AI model, QEV, achieving leading sensitive data detection performance compared with commercial AI models in data loss prevention (DLP), which helps enterprises prevent sensitive information from being leaked.

QEV analyzes input sentences to classify information and make decisions according to predefined criteria. While conventional generative AI focuses on answering questions or creating new content, QEV specializes in determining whether given information meets specific conditions and returning the appropriate result from a defined set of choices, together with the probability of each choice.

For example, when an employee enters customer information or unpublished business materials into an AI service, QEV can determine whether the text contains personal information or confidential company information. Enterprises can use these results to identify the risk of sensitive information being disclosed externally and to support review and response processes under their security policies.

QueryPie implemented its proprietary decision-making AI technology on Gemma 4, an open-weight AI model, and developed QEV specifically for sensitive data detection through additional training and optimization using enterprise security operations data. The company said the model is designed to support further training with each enterprise’s data and business requirements, enabling use across a range of enterprise environments.

QEV Achieves a 99.69% Sensitive Data Detection Rate

To validate QEV’s performance, QueryPie compared its internally developed model with an external commercial AI model in a DLP detection evaluation using internal documents in Korean, English, and Japanese.

QEV achieved a sensitive data detection rate of 99.69%, correctly identifying 1,277 of the 1,281 documents containing sensitive information. In the same evaluation, the external commercial decision-making model JEV API missed 46 documents containing sensitive information, while QEV missed only four, reducing missed detections by approximately 91.3% compared with the commercial model.

QEV also recorded an F2 score of 98.98%, a combined measure of sensitive data detection performance. The F2 score evaluates both precision—the accuracy of information identified as sensitive—and recall—the ability to detect sensitive information without missing it.

This result demonstrates the significance of developing an in-house model with state-of-the-art DLP performance by applying QueryPie’s proprietary technology and security-specific data to further train and optimize a publicly available AI model.

Sensitive Data Detection Performance: QEV vs. a Commercial AI Model

The following results compare both models under the same conditions using 1,281 documents containing sensitive information.

Metric	Commercial AI Model (JEV)	QEV
Detection rate	96.41%	99.69%
Successful detections	1,235	1,277
Missed detections	46	4

Context-Based Classification of Personal and Unpublished Business Information

QueryPie also conducted a demonstration test using QEV on 10 fictional emails and AI input examples representing situations that could arise in enterprise operations.

The test classified input text into four categories: no sensitive information, personal information, confidential company information, and both personal and confidential company information. Text containing customer names, contact details, and shipping addresses was classified as personal information, while text containing enterprise pricing and discount terms before their announcement was classified as confidential company information.

When a sentence included both customer information and unpublished business information, QEV identified both types of sensitive information. Public event announcements and ordinary everyday conversations were classified as containing no sensitive information. All 10 test cases returned results consistent with the predefined criteria.

QEV can determine whether sensitive information is present and distinguish its type, including personal and confidential company information, based on context. Enterprises can use these decisions to understand which information may violate security policies and support security processes such as sensitive information review and approval.

High Detection Performance and Operational Efficiency: Up to 45.2% Less GPU Memory

Operating AI models generally requires substantial computing resources and GPU memory. For enterprises running AI models on their own infrastructure, model memory requirements can significantly affect operating costs and infrastructure design.

To reduce this burden and improve operational efficiency, QueryPie applied INT8 quantization, which lowers the numerical precision used in model computations to reduce memory usage. A comparison of performance and GPU memory usage before and after quantization across the QEV model family showed a 27.5–45.2% reduction in GPU memory usage while broadly maintaining sensitive data detection performance.

In particular, the QEV model that achieved a 99.69% detection rate continued to miss only four documents after quantization, while its F2 score improved from 98.98% to 99.02%.

QEV therefore combines high sensitive data detection performance with greater operational efficiency through reduced GPU memory usage. QueryPie plans to continue optimizing model performance and operating efficiency for deployment in real enterprise environments.

Decision-Making AI That Can Be Specialized for Each Enterprise’s Domain

QEV is designed to make decisions required across a range of enterprise tasks, extending beyond sensitive data detection. QueryPie designed the model to support additional training and optimization for each enterprise’s security policies and decision criteria, and plans to expand its use to enterprise tasks such as information classification and condition-based decisions.

“QEV is a decision-making model developed to use AI to make the judgments required across diverse enterprise environments,” said a QueryPie spokesperson. “Building on this DLP performance validation, we will continue advancing AI technology that enterprises can use in real operations and expand its application beyond security to a wider range of enterprise domains.”

QueryPie has made its QEV model available on Hugging Face, an AI model sharing platform, so that anyone can directly evaluate its performance and potential applications.

The company also provides a separate demo page where users can try QEV’s decision-making capabilities. Available in Korean, English, and Japanese, the demo lets users view classification results and the probability of each choice using predefined test cases or their own text.

• Hugging Face: https://huggingface.co/querypieai
• Demo: https://qev.demo.querypie.com

About QueryPie

QueryPie AI is an enterprise security and AI company founded in Silicon Valley in 2016. It provides data and system access control solutions optimized for cloud and on-premises environments, as well as an AI Platform.

Officially launched in June 2020 through collaboration among engineers from Kakao and Naver and experts from various fields, the Access Control Platform (ACP) has expanded across access control and auditing for data, infrastructure, and business systems. More recently, QueryPie has added AI capabilities on top of its security foundation, helping enterprises use AI tools and AI agents while consistently maintaining permissions, approvals, auditing, and sensitive information protection.

With the launch of AI Platform (AIP) in 2025, QueryPie has expanded into connecting the SaaS applications, data, and internal systems used by enterprises with AI agents to enable real business automation and AI transformation (AX). AI Platform offers AI apps developed and operated directly by QueryPie AI, including Lingo, a real-time interpretation service, and NotePie, which makes it easy to turn internal company data into the materials users need.

QueryPie AI provides a proven security platform for enterprise customers, backed by global security certifications and compliance capabilities including SOC 2 Type II, CSA STAR, ISO/IEC 27001, ISO/IEC 42001, and ISMS-P. The company has raised a total of US$36 million from global investors including Y Combinator, Salesforce, and SoftBank-affiliated Z Venture Capital. With 80% of South Korea’s unicorn companies as customers, QueryPie AI is accelerating its international expansion, focusing on Japan, the United States, and Southeast Asia.

Media Contact

QueryPie PR Team

• Email: pr@querypie.com
• Website: https://www.querypie.com/