Selectstar unveils DAVI, a decision-specialized AI model

Selectstar, an AI reliability evaluation company, has released ‘DAVI’ (Decision AI), a specialized model designed to assess question difficulty and select the most appropriate AI model for a given task. This process, known as model routing, aims to balance AI service performance with operational costs.

In experimental testing involving 870 high-difficulty Korean Q&A instances, DAVI routed tasks between Anthropic’s Haiku, Sonnet, and Opus models. The routing achieved a 71.1% accuracy rate, slightly surpassing the 70.3% accuracy of using the high-performance Opus model exclusively, while reducing execution costs by approximately 28%.

DAVI also demonstrated superior performance in comparative benchmarks against TypeSafe AI’s ‘Jev’ model. On the JevBench evaluation, DAVI recorded 89.2% accuracy compared to Jev’s 86.6%. In Korean feedback evaluations, DAVI achieved 72.5% against Jev’s 66.8%.

Selectstar plans to integrate DAVI into its ‘Datumo Eval’ platform in November to expand AI quality evaluation capabilities. Additionally, the company has released a model routing plugin under the Apache License 2.0 on GitHub for open-source use.