Case Study: TypeSafe's Jev Model Beats Frontier LLMs on Classification Tasks

TypeSafe AI published a case study describing several classification tasks a customer measured after adopting Jev, TypeSafe's typed decision model built for closed-set choices: matching repeat analytics questions to approved answers, picking one metric out of 36, blocking requests containing personal information, and sorting expenses into roughly 55 categories.

TypeSafe reports Jev generally outperformed frontier LLMs on these tasks at a fraction of the cost: repeat-question matching accuracy rose from 70% to 97%, and expense categorization accuracy rose from 50% to 86% against human reviewers. Speed, measured by shadowing live production traffic, was up to four times faster than the frontier-model baseline, and up to three times faster in offline tests. These figures come from TypeSafe's own case study rather than an independent benchmark, so they describe what worked for one customer's specific tasks rather than a general guarantee; What Is Jev explains the decision-model design behind these numbers.