SelectStar unveils DAVI, the AI that chooses the right model and reduces execution costs

The Korean company says its decision-making model beats Jev in accuracy tests and automatically selects Anthropic models with a 28% cost saving; adoption is planned for November in the Datumo Eval platform.

What SelectStar announced

SelectStar has introduced DAVI, an artificial intelligence model dedicated to decision-making—that is, a system designed to analyze a situation and indicate the most appropriate response—according to a report by ETNews on October 26. The main function described is model selection: DAVI assesses the difficulty of a question and chooses the most suitable AI model, or judges the adequacy of an answer already produced. The company also points to a possible use in predicting the probability that an AI agent will successfully complete a task.

The numbers in the comparison with Jev

To validate performance, SelectStar compared DAVI with Jev, TypeSafeAI's model, in five areas. In the JevBench evaluation tool, DAVI reached an accuracy of 89.2% versus 86.6% for Jev. On high-difficulty problems, characterized by complex conditions and nuanced judgments, DAVI scored 77.5%, exceeding Jev's 73.0% by 4.5 percentage points.

The company also measured automatic model selection on 870 high-difficulty Korean questions and answers. DAVI chose among Anthropic's Haiku, Sonnet and Opus models based on question difficulty, achieving a correct-answer rate of 71.1%, higher than the 70.3% obtained using Opus alone. In the same test, according to the company, the model's execution cost fell by about 28%.

Adoption plan and openings

SelectStar plans to integrate DAVI in November into its AI reliability evaluation platform, Datumo Eval, to enable companies to verify the results of automated decisions and improve operational efficiency. The company has also released as open source, under the Apache 2.0 license, a DAVI-based plugin for automatic model selection.

The material also reports a previous agreement: a letter of intent with Cisco to collaborate on AI security and reliability, combining Cisco's AI Defense platform with SelectStar's evaluation technologies, with the aim of exploring verification methods tailored to companies. It remains to be clarified how these solutions will actually be integrated and which customers will adopt them.

Source and attribution: this original summary was prepared by AICronaca based on the report by ETNews — AI ↗. The statements are attributed to the source; any information limitations: The text is based on a press release reported by ETNews and on data provided by the company; the comparative results have not been independently verified. The material does not specify complete methodologies, sample sizes beyond the 870 questions cited, or details on absolute costs. The news about OpenAI and the three AI safety scenarios appears in the material without context and has not been included..