[AI Brief] SelectStar’s Financial AI Safety Evaluation Work Recognized at Data Mining Conference

IT Daily provides an "AI Brief" section that brings together major AI news from Korea and abroad.

In one item in the section, SelectStar said on the 5th that its research paper on financial AI safety evaluation had been accepted to the main conference of "IEEE ICDM 2026." "IEEE ICDM 2026" is a globally prestigious conference in the data mining field.

The paper was selected as an oral presentation paper, accounting for about 8.9% of all submissions.

The photo accompanying the article introduces the technical architecture of the financial AI safety evaluation framework "FinRED." The diagram shows that "FinRED" includes financial risk scenario generation, expert-guided evaluation criteria establishment, and an AI model safety verification process. The photo is courtesy of SelectStar.

The accepted paper, co-authored by SelectStar AI researchers, the Korea Financial Security Institute, and KAIST, is titled "FinRED: An Expert-Guided Benchmark Generation and Evaluation Framework for Financial LLM Red-Teaming." The study presents a systematic evaluation technology for financial-specific LLM safety and proposes a framework to assess the safety of LLMs specialized for the financial sector.

To this end, the researchers built the "FinRED" framework, which generates risk scenarios by combining financial security expert knowledge and real financial documents. FinRED is designed to use financial expert judgment as an element reflected in the evaluation criteria, with the aim of assessing whether AI-generated responses support real-world risky behavior and to comprehensively determine the harmfulness of AI responses and whether they support risky behavior.

SelectStar said the study also demonstrated strong results in evaluating the risk of AI-generated responses. FinRED’s financial-specific evaluation criteria improved the match rate with financial security experts’ judgments compared with existing general-purpose evaluation methods. The match rate rose from 76.92% to 88.46%, an increase of 11.54 percentage points.

In addition, cases in which risky responses were mistakenly judged as safe decreased by about 57% compared with the previous approach. As a result, FinRED’s financial-specific evaluation criteria were found to be closer to financial security experts’ judgments than the general-purpose method, while also reducing misjudgments of risky responses.

Korean Deep Learning announced on the 5th that it had carried out a proof of concept, or PoC, related to financial operations with Shinhan Vietnam Finance (SVFC). The project was conducted through "Shinhan Futures Lab Vietnam," Shinhan Financial Group’s global startup nurturing platform.

The program is a global open innovation project hosted and organized by Shinhan Bank Vietnam and the Seoul Business Agency (SBA). It was set up to support collaboration and proof-of-concept projects between domestic startups and overseas financial institutions based on Shinhan Vietnam Finance’s technology needs.

Korean Deep Learning said it carried out a financial operations automation PoC with Shinhan Vietnam Finance, a Vietnam subsidiary of Shinhan Card. The purpose of the PoC was to verify the applicability of Korean Deep Learning’s technology. The demonstration applied Korean Deep Learning’s "Deep Agent."

"Deep Agent" is built on the AI information recognition technology "Deep OCR" and the structural analysis technology "Deep Parser." The solution can extract and structure needed information from a wide range of documents and link the extracted and structured information with actual work processes. In this demonstration, the company examined the applicability of information processing, data utilization, and follow-up work automation technology for local financial operations in Vietnam.

Based on this achievement, Korean Deep Learning plans to expand business opportunities in overseas financial markets, including Vietnam. To that end, it plans to leverage the technology and build experience it has accumulated with domestic companies and public institutions, and to explore ways to apply solutions suited to local financial institutions’ work environments and operational needs. It also plans to expand opportunities for cooperation with overseas partners.

(Photo: Infozis)

Infozis announced on the 5th that it will take part in "Bixpo 2026," an international electric power and energy technology exhibition. Infozis plans to unveil the results of its demonstration project for an "artificial intelligence-based virtual power grid" that it is pursuing with Korea Electric Power Corporation.

In power grid operations, it is necessary to comprehensively review the complex connections among transmission and distribution facilities and operational criteria. In this process, it has been raised as an issue that the same facility is named differently across systems and that connection information between facilities is scattered across multiple documents.

As a result, if the connection relationships are misinterpreted, there is a possibility that facilities subject to review will be omitted. Conversely, facilities that are not actually connected may be included in the analysis.

To address these issues, Infozis participated in KEPCO’s national power grid AI advancement project. Ontology-based data modeling technology was applied to the project.

Infozis structured the connection relationships among power facilities and the operating and technical criteria into a single semantic framework. Based on this, it designed the AI to query and analyze information according to actual facility relationships and review criteria.

Infozis is the contractor for KEPCO’s AI-based virtual power grid demonstration project. In this project, Infozis was responsible for ontology-based power grid data structuring and AI agent development. The project is scheduled to wrap up in November.

Tak Jeong-su, CEO of Infozis, said the company plans to expand the application of ontology-based enterprise AI. The targets for expanded application are major industrial sectors that require advanced data processing capabilities and high accuracy.

Yejin unveiled "Jelly Law Assistant," a legal information service based on its own Korean model. Yejin announced on the 5th that it would provide "Jelly Law Assistant" free of charge. The model behind "Jelly Law Assistant" is Yejin’s in-house Korean foundation model "Yejin 8B v3." "Jelly Law Assistant" can search related legal provisions using only natural language questions.

Yejin said that "Yejin 8B v3" did not use weights from overseas open models. It also said that "Yejin 8B v3" is a From-Scratch pretrained model.

Jelly Law Assistant is a service aimed at the general public and working professionals who are not familiar with legal terminology. When a user enters a natural-language situation, it uses a method that combines RAG and a law-specialized sLM to organize and provide relevant current legal provisions and the basis for them.

According to Yejin, the service was designed to suppress hallucinations. To that end, it applies a fixed generation temperature of 0 and adopts a direct-answer No-Think design within the scope of retrieved provisions.

The base model is Yejin 8B v3. The model was developed in a single-node environment with 8 NVIDIA H200 GPUs and trained on 83.9 billion tokens.

The tokenizer is a self-designed Korean numeric-segmentation type with a vocabulary size of 64,000. The company claims that, through this tokenizer, it can reduce the number of tokens needed to process Korean legal documents by 24% to 38% compared with Llama 3 and others. The photo credit is Difly. The article also mentions that Difly supplied its sound analysis AI solution "Listen AI" to a global port operator.

Difly supplied its industrial sound analysis AI solution "Listen AI" to a global major port operator. The port operator is running port facilities in Busan, and Difly announced on the 5th that it plans to carry out sound-based predictive maintenance for automated crane equipment.

Listen AI is a solution used for quality inspection of industrial parts, inspection of connector and bolt fastening sounds, and predictive maintenance. The solution can accurately extract target sounds even in environments with noise above 80 decibels (dB), and it applies a technique that distinguishes abnormal sounds from normal sounds. The applicable sectors are mobility, home appliances, ports, robotics, and equipment.

Listen AI has been deployed at one terminal in Busan New Port operated by the global port operator. The target is an unmanned automated yard crane, and the predictive maintenance target is the hoist drive unit. Acoustic sensors have been installed on two motors and on the reducer, respectively.

The operating method is 24-hour continuous recording. About 300 hoisting cycle segments are automatically identified and analyzed each day, and when an abnormal sound occurs, alerts are sent to the terminal CMS and the remote office to prompt action.

Source: IT DAILY · Yang Seung-gap
Original: https://www.itdaily.kr/news/articleView.html?idxno=242708

References

• https://www.itdaily.kr/news/articleView.html?idxno=242708