Graph Technology Breaks Through the 70% Accuracy Barrier

Between 2024 and 2025, enterprises attempted to solve generative AI issues using vector technology. However, AI hallucinations still haven't disappeared, and accuracy remains at around 70%. The Infosys CEO explained, "In the early stages of generative AI adoption, we tried to solve these issues through vector technology. Even though it became hugely popular between 2024 and 2025 and was applied to enterprises, the hallucination problem in generative AI remains unsolved, and accuracy has stalled at around 70%."

The root cause is data fragmentation in enterprises. Companies have data scattered across departments and systems, and this dispersed information doesn't enable proper search functionality. Because AI lacks sufficient or systematically connected information to reference, it produces inaccurate answers.

Through this systematically connected data, AI can generate answers with more accurate evidence.

Infosys is conducting graph technology projects not in a single industry but across multiple sectors. In semiconductors, the company is pursuing projects that use graph technology for ultra-precise detection of security issues and optimization of production planning. The CEO explained, "Infosys is working on projects across diverse industry sectors including construction, plants, semiconductors, finance, and public institutions. Recently, we're conducting projects involving ultra-precise detection of security issues that could occur in semiconductors using graph technology, optimization methods for semiconductor production planning, structuring information for top national agencies, and applying blueprint recognition technology in petrochemical and energy sectors."

In the public sector, the company is structuring information for top national agencies, and in the petrochemical and energy sectors, it's applying technology to automatically recognize complex blueprints.

Infosys faced a technology sovereignty issue during its growth. The company won a project, but at contract time, the client demanded complete ownership transfer of the solution. The CEO recalled, "When we won the contract and reached the signing stage, the company demanded that we transfer all ownership of the solution to them. (omitted) We ultimately decided not to sign that contract. The reason was that while we would receive project costs initially, it would be extremely difficult to maintain sovereignty over our technology itself, so we abandoned the project."

Despite the financial impact, this decision became the turning point for establishing the company's foundation. By not surrendering the technology, the company could continue developing its independent platform.

There was a moment when Infosys's graph technology proved its true value. A client had been searching for over six months for a way to quickly recognize around 3,000 blueprint drawings, and Infosys proposed using a graph database. The CEO explained, "They received about 3,000 blueprint drawings and had been searching for over six months for a way to quickly recognize them. I suggested, 'Why don't we try reading them with a graph database?'"

Although Infosys didn't have vision technology, the company researched all relevant papers and vision recognition attempts in the field and created a solution with the team. The CEO said, "We didn't have any vision technology. We started by saying we'd give it a try, and we went through all the papers related to the domain and what attempts had been made in vision recognition, creating something with the team at that time. I thought it might not work, but when we actually tried it, it worked." This achievement secured global-level technological capability and earned the trust of large corporations.

Specifically, the company is creating an AI feature to companywide gather how FDEs (forward-deployed engineers) approached various concerns in each project and why they made certain choices. Eun-Yoon said, "We're creating an AI feature to collect across the company how FDEs approached their work, what concerns they had, and why they made certain choices. As we dive deeper into companies' concerns in the future, we're establishing an axis to capitalize on these experiences so our starting point can be sufficiently far ahead." The accumulated experience becomes the starting point for new projects, accelerating development speed.

This project is being proactively pursued by clients and is considered an important inflection point for the company because it's technology development closer to the physical world. Eun-Yoon said, "Since clients are proactively pursuing this project, I see it as an important inflection point right now, as it's a project closer to the physical world."

Technology lead Jin Chang-ha explained the differences and processing methods for structured and unstructured data. Structured data (Excel, databases) is easy for machines to handle, but unstructured data (emails, contracts, reports) has complex content and context, making it very difficult to process. Jin Chang-ha said, "We go one step further—we extract what concepts the structured documents are dealing with and what their subjects are."

Infosys introduced an 'ontology layer' that goes beyond the limitations of vector RAG methods. This simplifies and abstracts complex concepts, expressing data relationships visually as if buildings exist and roads connect them. Jin Chang-ha explained, "We call it an ontology layer—buildings exist and roads connect them in this way. It's a concept that simplifies and abstracts complex concepts." When concepts and relationships are abstracted this way, then paths are found and answers derived, accuracy improves dramatically.

In projects where accuracy is critically important, the strategy is to intentionally lower AI's degrees of freedom. Jin Chang-ha explained, "I realized that the fewer degrees of freedom we give AI, the higher the accuracy becomes. Because more degrees of freedom means more room for mistakes. When reducing degrees of freedom, DSL could be beneficial."

Infosys improved performance by implementing DSL (Domain-Specific Language) on its platform. DSL is a restricted language used only in specific industries or problem domains, enabling AI to operate only within defined parameters. Simultaneously, to improve data quality refinement, field specialists and engineers collaborate closely to enhance the quality of input values the machine learning model receives.

Infosys's FDEs (Forward-Deployed Engineers) differ from typical SEs (Solution Engineers). FDEs don't simply provide technical support—they're technical experts who directly enter customer sites, uncover, and solve customer problems. FDE Shin Ji-on explained, "We're called forward-deployed engineers, and we're engineers who directly enter customer sites to solve their problems."

Shin Ji-on performs work that directly contributes to corporate operating profit and stock prices through a product simulation agent. Additionally, by introducing OQL (OQL), an intermediate layer that enables LLMs to understand language more easily, the company reduced hallucination rates to nearly 0%. Shin Ji-on said, "We created an intermediate layer called OQL to make it easy for both us and LLMs to understand, which allowed us to reduce hallucination rates to nearly zero."

Infosys's next-generation security technology follows a completely different paradigm from existing security products. Traditional approaches register predefined attack patterns in a database and block only matching threats. Conversely, Infosys's product doesn't presume such rules from the start. ML lead Pyeon Myeong-jang explained, "Our product follows a completely different paradigm from existing security products. We don't presume such rules initially."

The new technology learns normal business patterns first, then detects anomalies that deviate from those patterns. Some attacks remain dormant for 30 days while exfiltrating data, which can only be detected by analyzing long-term logs in graph form. Pyeon Myeong-jang explained, "Some attacks remain dormant for 30 days while sending data outside, and these attack patterns can only be caught by recognizing long-term logs in graph form." Infosys was the first in Korea to demonstrate this graph-based anomaly detection technology in a large-scale enterprise environment.

Automatically restoring blueprint connection relationships was a difficult unsolved problem even in academia. Most companies had experts manually analyze blueprints. ML lead Pyeon Myeong-jang explained, "This was mostly done manually by people. When done manually, it takes a minimum of 8 hours."

Using Infosys's technology for automatic recognition reduces this process to under 5 minutes. Pyeon Myeong-jang said, "When we use our product for automatic recognition, this comes in under 5 minutes, significantly improving actual work efficiency." Currently, Infosys is conducting joint research with Seoul National University on this technology.

This demonstrates that Infosys prioritizes technology value creation over revenue maximization.

One Infosys team member reached their current position through a diverse background. They worked as an astronomer in Taiwan, achieving several significant results that were published in international academic journals. Later, they came to Korea to pursue a master's degree in data science at Seoul National University while currently developing AI agents at Infosys. They introduced themselves: "While working as an astronomer in Taiwan, I completed several important achievements and published them in international academic journals. I'm currently developing AI agents at Infosys."

This demonstrates that Infosys is attracting talent from diverse backgrounds including data science, machine learning, and domain expertise.

Infosys's technological achievements are validated through customer reactions. A major energy company client expressed surprise exceeding expectations upon witnessing the current technology level's accuracy. This isn't mere praise—it signifies that accuracy levels previously thought impossible in the industry have now been achieved.

Authenticity emerges from how the Infosys CEO talks about the problems being solved. The CEO naturally says "the problems we solve are fun," suggesting commitment to solving technical problems themselves rather than merely pursuing business success. The CEO innocently asks, "How could solving problems not be fun?"—reflecting the company's overall culture and values.

Infosys's competitiveness rests on two core technologies. Blueprint recognition technology is an area with no industry precedent—the ultimate frontier. The company also possesses clear strengths in vision technology. Additionally, it holds the most experience in creating ontology from unstructured data.

The combination of these two technologies creates highly sophisticated competitiveness that competitors find difficult to match. No other company besides Infosys can currently simultaneously process unstructured visual information like blueprints and unstructured text data while expressing their relationships through ontology.

Infosys team members emphasize Korea's AI self-sufficiency. They advocate leveraging advanced technologies like Palantir and Nvidia while building Korea's own independent AI infrastructure. A team member argued, "We need to use advanced technologies that enter our country while building AI infrastructure that represents our own sovereignty."

The key point is that Korea needs capability not to export data but to export the methodology itself. The team member emphasized, "It's an extremely important technology to export the methodology itself, not data." This suggests that AI-era technology independence is a core task for Korea to escape technological dependence.