SelectStar, an AI reliability evaluation company, said on Oct. 26 it unveiled DAVI, a decision-focused AI model.

DAVI is a model specialized in decision-making tasks, such as selecting an appropriate AI model depending on the difficulty of a question or evaluating whether an AI answer is appropriate. It can also be used to predict the likelihood of an AI agent succeeding in its task.

SelectStar said it compared DAVI's performance with Jev, a global decision AI model from TypeSafeAI, and found DAVI recorded higher accuracy in major evaluation items.

On 231 questions in JevBench, which evaluates AI decision-making capability, DAVI posted 89.2 percent accuracy, beating Jev's 86.6 percent by 2.6 percentage points. On high-difficulty questions, DAVI scored 77.5 percent, 4.5 percentage points higher than Jev's 73.0 percent. On Korean-language feedback evaluation, DAVI also outperformed Jev, scoring 72.5 percent versus 66.8 percent.

SelectStar also confirmed cost savings in routing experiments to select AI models. It had DAVI choose the most suitable model among Anthropic's Haiku, Sonnet and Opus models for 870 Korean-language high-difficulty Q&A cases.

DAVI-based routing delivered a 71.1 percent correct-answer rate, 0.8 percentage points higher than when using Opus alone at 70.3 percent. Model execution costs fell by about 28 percent compared with using Opus alone.

SelectStar plans to apply DAVI in November to its AI reliability evaluation platform, Datumo Eval. It said it aims to support companies' AI quality evaluation and more efficient model operations.