Flitto CEO Lee Jung-soo: “Staying Half a Step Ahead Is a Survival Strategy”

Flitto is expanding its business beyond supplying AI training data into data distribution and translation and interpretation solutions. Flitto CEO Lee Jung-soo unveiled this expansion strategy in an interview with Techworld on Sept. 30. The company is building on its data processing capabilities as the basis for expansion. Flitto has accumulated multilingual data through its crowdsourced translation platform.

Flitto plans to add a marketplace to its existing customer-tailored data build business and introduce an integrated brand, Vogl. It is also pushing ahead with a plan to use enterprise on-premises systems. The goal is to broaden its revenue base. The idea is to process materials held by multiple suppliers for different training purposes and then sell them. It is also pursuing a plan to expand usage by personalizing translation and interpretation services and linking functions.

This expansion is an extension of a business that began with securing data for machine translation training. CEO Lee said the platform’s original purpose from the beginning was to secure data for machine translation training. Flitto has built text, then voice, image, and video data, and recently expanded its business scope to data for physical AI training. Lee presented staying “half a step ahead” as a survival strategy.

Flitto operates a structure in its data marketplace that receives registrations of materials held by companies and individuals and supports processing and sales. Rather than simply accepting material registrations, it is broadening its role by taking on processing and sales as well. As a differentiating factor, the company cited improving commercial value through processing rather than acting as a simple intermediary.

As a result, the scope of the business is also expanding. Flitto is shifting from a model centered on customer-specific data builds to one that uses materials from multiple suppliers and responds to a wider range of purchasing demand. CEO Lee explained that Flitto is participating in the original-data processing segment.

Lee said integrating data posted by multiple companies can produce better data. A representative example was articles held by news organizations. He explained that articles owned by news outlets can be sold as AI training data.

However, the time and cost involved in processing vary by data type. In particular, CoT data and code data require work and review by specialized personnel. The processing burden differs depending on the type of data.

Code data must be written by a developer and then verified by another developer. For that reason, code data processing requires substantial manpower and time. Types such as CoT and code that require specialized review were presented as areas where time and costs are significant.

Lee said that to produce high-quality data, both the writer and reviewer must be skilled developers. He explained that because the input of labor and time is large, data sales prices are also high.

He said that in the data market, speed in securing data used to be the top priority, but recently, as customer quality demands have become more sophisticated, expertise that meets each customer’s verification standards has become an important factor.

Flitto said it is improving accuracy by combining specialized human work with AI tools. Lee said humans handle the initial work, while AI tools are used in the process of improving accuracy. He added that the company is meeting the quality demanded by customers by using its own technology together with AI.

Lee pointed to the National Institute of Korean Language corpus build project as an example. He said the project was pursued in the context of linking Korean with low-resource languages, and that it was carried out by assigning dedicated staff and then conducting verification to accurately reflect the Korean language and cultural context.

Flitto is using its marketplace as a channel to introduce data from various regions and fields to overseas customers. Lee said some customers still assume Korean companies deal only in Asian data, and that the company intends to change that perception through the data it actually supplies.

The overseas share of data sales is also high in actual revenue. According to Flitto’s semiannual report, data sales revenue in the first half of this year was about KRW 6.594 billion, and data sales exports in the first half were about KRW 5.579 billion. The export share was 84.6%.

At the same time, Flitto is reshaping its solutions business around Vogl. Vogl was recently unveiled at a press briefing and is a brand that encompasses multiple services. The brand covers translation and interpretation services for personal conversations and meetings, as well as “Vogl Stage” for events and lectures and “Vogl Zone” for stores and counters.

Vogl also includes enterprise internal build on-premises solutions. Flitto plans to link its interpretation and recording functions and allow voices and terminology learned for each user to be used across multiple services.

Lee said that in the past, there was confusion because users who saw ads for the translation and interpretation solution would download the Flitto app, which was for data collection. He added that introducing the Vogl brand has made it possible to clearly distinguish and deliver the services.

Flitto is emphasizing its own model training and operation capabilities in its on-premises business. Because on-premises translation and interpretation must be provided within a customer’s internal environment, the company sees a need to directly train and process models and maintain performance even with limited server resources. Lee explained that supplying on-premises solutions to enterprises with high security requirements requires the ability to directly process and train models. He also cited model lightweighting and maintaining accuracy and speech recognition rates as examples of the company’s technological strength.

On-premises contracts are characterized by their large scale. However, on-premises systems also carry the burden of requiring substantial manpower and cost at the initial build stage. Lee said that as the number of customers increases, scale economies can be secured by having existing staff manage multiple systems.

Based on its domestic build experience, Flitto plans to pursue entry into the Japanese and Middle Eastern markets. The company aims to connect the build experience accumulated in its on-premises business to overseas expansion.

Flitto sees personalization as a factor that encourages continued use by customers. The more a user’s voice and terminology are learned, the more advanced the translation and interpretation becomes for that customer. This personalization effect was presented as a factor that reduces the need to switch to other services.

Lee said that as personalized engines are trained more, users have less need to move to other services. He explained that the effect of such personalization creates a virtuous cycle for the company.

Building on this foundation, Flitto is expanding into data for physical AI training. The company is collecting videos of human behavior through crowdsourcing.

The target data range for the build is for training on task success, failure, and recovery after failure. Flitto already has an existing user base and also has experience in data cleansing.

Lee said the company can secure collection speed based on its existing user base and data-cleansing experience. He added that the same foundation can also secure quality.

Lee said competitors are chasing with similar technology or lower prices. He stressed the need to preemptively respond to new data demand. Accordingly, the company plans to continuously improve its data collection and processing methods. As a survival strategy, Lee said the company should always stay half a step ahead of its rivals, and that if competitors catch up, it will move half a step ahead again.

On the policy side, he suggested that the elements needed to advance AI should not be limited to infrastructure and model development, but should also include securing training data. Accordingly, when designing a physical AI business, the costs of data collection and processing need to be reflected, and the participation of specialized companies also needs to be reflected, he said. He also noted that in the process of using overseas data, including data from China, it is necessary to reflect the Korean language context and Korean cultural context.

At the company level, Lee said the goal for next year is to establish the data marketplace and generate meaningful revenue next year. He added that the company plans to actively promote solutions under the Vogl brand and to carry out marketing with a focus on return on investment.

Flitto plans to push ahead with expanding the data marketplace in 2027, popularizing Vogl-based solutions in 2027, and building customer-tailored on-premises systems in 2027. At the same time, Flitto plans to continue its existing data supply business, expand marketplace transactions, grow overseas solution customers, and broaden its revenue base.

Source: TECHWORLD · Kim Seung-gi
Original: https://www.epnc.co.kr/news/articleView.html?idxno=407690

References

• https://www.epnc.co.kr/news/articleView.html?idxno=407690