Samsung Develops Agent Builder, SIRIUS, and Reference Image-Based AI

Samsung aims to make AI a core part of its internal operations by developing in-house models and tools. The company’s proprietary large language model, Gauss, is playing a big role in this effort. It has recently developed tools like Agent Builder, SIRIUS, and a reference image-based image generation model.

Samsung strengthens internal AI capabilities with proprietary models

According to a report from The Elec, Samsung Research has developed a customized agentic AI production tool called Agent Builder. The tool uses a bunch of LLM models, including Gauss 2.3, Gauss 2.3 Think, and Gauss O Flash. It helps to create AI agents by combining different components. One key component is DoXA, which helps AI agents understand business-specific documents. Furthermore, Agent Builder offers a drag-and-drop interface so employees do not need to write any code.

Apart from Agent Builder, Samsung Research has developed SIRIUS, a knowledge search system powered by Gauss. Unlike traditional search methods, SIRIUS focuses on graph-based information structuring. “Sirius is currently being provided as a beta service for employees,” said an official from Samsung Research’s Software Innovation Center. “It is mainly used for searching for product development-related knowledge or exploring specific technical tasks and information on those in charge.”

Samsung has also created a reference image-based image generation model. It accepts both natural-language prompts and reference images. Existing image generation methods cannot create accurate images only from text, and struggle to generate images for unlearned knowledge. Now the upgraded model lets the AI use visual information from the reference to create new images.

The company saw model usage increase by 153% per month after its internal release. However, developing the model was anything but without challenges. Samsung researchers needed foundation models that were open-source and commercially usable. However, many high-performance models prevent additional learning due to technical limitations. To solve this, the team built specialized datasets and optimized the models.