Jev in Practice: Browser Use and Agent Workflows

The author built a Jev dashboard that automatically labels 1,315 posts across 8 dimensions such as topic, hook and writing style for about $0.086 in estimated model cost, then filters by tag and compares engagement.

An explanation of Jev in LLM terms: it does no text generation and only returns three structured decision types for a given state — Noul, Choice and Score.

A developer plans to build an open-source Jev for a YouTube project, discussing a bidirectional attention transformer, masked-LM training and fixed-schema structured output, and comparing three output-head designs.

Bespoke trained Nimble with a light LoRA on 2,676 contrastive examples, reaching 90.12% on its 324-example held-out set versus Jev's 93.21%, reading logits directly without generating CoT.

In an early stll_app demo, an LLM reads the house style guide once, and at runtime Jev picks the right style for each paragraph based on the document's existing structure.

The author uses Jev to make decisions in Tetris, reaching the final level 30 without pausing, to show how fast it makes choices.

A video walks through TypeSafe's Jev step by step: you send state text with predefined options, and it returns an answer plus a probability, targeting the cost of yes/no decisions inside agents.

A developer says using Jev to make decisions before sending an email generated about $1,250 in extra sales in one night, paying back the API cost for a year or two.

The author argues that demo wrappers giving Jev context and options aren't fake, since Jev still makes the decisions, and encourages people to build freely with it.

Developer chadfowler asks for the best Jev demo so far, and the replies collect various examples of how people are using it.

The author says they did nothing themselves: Amp built the whole thing and recorded the video inside an orb, fully autonomously, using Jev to create a Copilot-like tool.

The author clarifies that Jev is a fast classifier, scorer and decision maker that doesn't generate free-form text or code, making it a companion to LLMs rather than a replacement.