The Finnish startup creates Root Judge, an innovative AI model evaluation tool trained on Europe’s largest supercomputer.
As generative artificial intelligence becomes a cross-sector technology in areas such as education, healthcare, public administration, and industry, doubts about its reliability, transparency, and control are also growing. In this context, the startup Root Signals has achieved a notable breakthrough: the development of Root Judge, an AI evaluation model specifically designed to assess the accuracy of other large language models (LLMs).
The most significant aspect is that this advancement was made possible through the use of the LUMI supercomputer, the most powerful in Europe and one of the fastest and most sustainable in the world, located in Kajaani, Finland. The project has benefited from the support of Business Finland and access to resources from the EuroHPC program, a clear demonstration of the potential for public-private collaboration in developing strategic technologies.
A “judge” model trained with 400 GPUs
Root Judge was conceived as a model capable of detecting hallucinations (generative errors) in other LLMs and transparently justifying its evaluations. To train it, the Root Signals team utilized nearly 400 LUMI GPUs, processing vast volumes of synthetic and open-source data. The result is a flexible and open-source model that outperforms leading market models in evaluation tests, including those from OpenAI and Anthropic.
“We wanted to demonstrate that it is possible to evaluate the quality of an AI model with another specific model designed for that purpose, and to do so transparently,” explains Oguzhan Gencoglu, project director and primary user of the LUMI system.
To facilitate adoption, the team opted to quantify the model and reduce its computational needs, making it viable for companies that do not have large hardware infrastructures.
LUMI as a technological catalyst
Root Signals was founded in 2023 in Helsinki, and after receiving pre-seed investment and institutional support, it gained access to LUMI’s advanced computing resources. The team leveraged technical support from the CSC – IT Center for Science to overcome initial barriers and begin training the model.
However, they acknowledge that development timelines do not always align with funding schedules. “Greater flexibility in project durations would allow better utilization of available resources and avoid time constraints that could affect development,” Gencoglu points out.
Growth, reputation, and open-source strategy
Since the launch of Root Judge, the company has gained traction in the market, acquiring new clients in sectors that need to ensure the reliability of their AI-based tools. The model is offered as open-source and free for commercial use, which has strengthened Root Signals’ image as a reliable provider in a sector dominated by large tech firms.
“We wanted to build trust. Publishing Root Judge as open source was key to showing what we are capable of,” adds Gencoglu. This strategic decision has not only fueled growth but also established the startup as a benchmark player in quality control for AI-based applications.
The next challenge: AI with reasoning capabilities
Looking ahead, Root Signals is not content with the success of Root Judge. Its next ambitious goal is to develop an AI model with reasoning capability, able to not only assess but also explain complex decisions in a human-like manner.
“No one has yet trained an evaluator model with integrated reasoning. It’s a significant technical challenge, but we believe it’s the next step in responsible AI evolution,” assures Gencoglu.
With projects like Root Judge, Finland and Europe are reinforcing their commitment to a more transparent, reliable, and accessible artificial intelligence, supported by infrastructures like LUMI and a tech community committed to open source. In an increasingly automated world, tools like this will make the difference between blind adoption and informed trust.
Source: AI News