Seoul-based Datumo Raises $15.5m To Take On Scale Ai, Backed By Salesforce

Most organizations opportunity they aren’t afloat prepared to usage generative AI successful a safe and responsible way, according to a caller McKinsey report. One interest is explainability – knowing really and why AI makes definite decisions. While 40% of respondents position it arsenic a important risk, only 17% are actively addressing it, per nan report.

Seoul-based Datumo began arsenic an AI information labeling institution and now wants to thief businesses build safer AI pinch devices and information that alteration testing, monitoring, and improving their models—without requiring method expertise. On Monday nan startup raised $15.5 million, which brings its full raised to astir $28 million, from investors including Salesforce Ventures, KB Investment, and SBI Investment, among others.

David Kim, CEO of Datumo and a erstwhile AI interrogator astatine Korea’s Agency for Defence Development, was disappointment by nan time-consuming quality of information labeling truthful he came up pinch a caller idea: a reward-based app that lets anyone explanation information successful their spare clip and gain money. The startup validated nan thought astatine a startup title astatine KAIST (Korea Advanced Institute of Science and Technology). Kim co-founded Datumo, formerly known arsenic SelectStar, alongside 5 KAIST alumni successful 2018.

Even earlier nan app was afloat built, Datumo secured tens of thousands of dollars successful pre-contract income during nan customer find shape of nan competition, mostly from KAIST alumni-led businesses and startups.

In its first year, nan startup surpassed $1 cardinal successful gross and secured respective cardinal contracts. Today, nan startup counts awesome Korean companies for illustration Samsung, Samsung SDS, LG Electronics, LG CNS, Hyundai, Naver, and Seoul-based telecom elephantine SK Telecom among its clients. Several years ago, however, clients began asking nan institution to spell beyond elemental information labeling. The seven-year-old startup now has much than 300 clients successful South Korea and generated astir $6 cardinal successful gross successful 2024.

“They wanted america to people their AI exemplary outputs aliases comparison them to different outputs,” Michael Hwang, co-founder of Datumo, told TechCrunch. “That’s erstwhile we realized: we were already doing AI exemplary information — without moreover knowing it.” Datumo doubled down connected this area and released Korea’s first benchmark dataset focused connected AI spot and safety, Hwang added.

“We started successful information annotation, past expanded into pretraining datasets and information arsenic nan LLM ecosystem matured,” Kim told TechCrunch.

Meta’s caller $14.3 cardinal acquisition-like finance successful data-labeling institution Scale AI highlights nan value of this market. Shortly aft that deal, AI exemplary shaper and Meta competitor OpenAI stopped utilizing Scale AI’s services. The Meta woody besides signals that title for AI training information is intensifying.

Datumo shares immoderate similarities pinch companies for illustration Scale AI successful pretraining dataset provisioning, and pinch Galileo and Arize AI successful AI information and monitoring. However, it differentiates itself done its licensed datasets, peculiarly information crawled from published books, which nan institution says offers rich | system quality reasoning but is notoriously difficult to clean, according to CEO Kim.

Unlike its peers, Datumo besides offers a full-stack information level called Datumo Eval, which automatically generates trial information and evaluations to cheque for unsafe, biased aliases incorrect responses without nan request for manual scripting, Kim added. The signature merchandise is simply a no-code information instrumentality designed for non-developers for illustration those connected policy, spot and safety, and compliance teams.

When asked astir attracting investors for illustration Salesforce Ventures, Kim explained that nan startup had antecedently hosted a fireside chat pinch Andrew Ng, laminitis of DeepLearning.AI, astatine an arena successful South Korea. After nan event, Kim shared nan convention connected LinkedIn, which caught nan attraction of Salesforce Ventures. Following respective meetings and Zoom calls, nan investors extended a soft commitment. The full backing process took astir 8 months, Hwang said.

The caller backing will beryllium utilized to accelerate R&D efforts, peculiarly successful processing automated information devices for endeavor AI, and to standard world go-to-market operations crossed South Korea, Japan, and nan U.S. The startup, which has 150 labor successful Seoul, besides established a beingness successful Silicon Valley successful March.