Jev in Practice: Browser Use and Agent Workflows

The author expands Jev's output vocabulary to 50k words using vocabulary sharding, hierarchical routing, and beam-search pruning, and builds a web page for chatting with it.

A demo shows Jev scoring branches and killing weak ones, with the score deciding which model runs, which tool is called, and which branch retries, claimed to be up to 193x faster and 444x cheaper in tests.

From Browser Use's click decisions and Claude Code context compaction to a Monad market-making bot and Postgres extensions, a list of ten repos that use Jev for decisions, with their star counts and purposes.

A workflow demo reads the goal and evidence, uses Jev to choose the next worker from fixed options, lets that worker research, write, or review, and loops while recording results, with a human approving publication.

Alan AI built an in-house review console where Jev checked 3,518 applications in 4:54, surfaced 20 for human review, at an estimated cost of $0.41.

A livestream demo uses opencode as the harness, Jev for decisions, DeepSeek 4.1F for reasoning and Chrome for execution, looping to find why the game ended and retune for a higher score.

The author recreated miu21590's Jev-driven effort router, which adjusts Astra's reasoning effort each turn to balance cost, performance and latency.

A hands-on look at Jev, a model that only handles classification tasks like yes/no, pick-one or scoring, skipping text generation to run 20-200x faster and cost 40-400x less than a normal LLM.

The author uses the Jev classifier to score each step from 0 to 1, letting Claude Opus 5 switch dynamically between LOW, MEDIUM and HIGH reasoning, finding 6 bugs in 14 steps for $0.38.

The author plugs a Twitter narrative analyzer into Nerve that outputs accelerating narratives every 3 minutes, which Jev then matches against live pools within 190ms, claiming it caught an AI agent sector rotation 47 minutes early for +0.9 ETH.

The author turns casestudies.com's 6,779 B2B categories into a scoring queue and uses Jev to score each one against a fixed set of questions such as ACV, sales cycle and buyer universe size, ranking them for research priority.

A free legal citation checker built with Jev that scans a 7-page memo in under 9 seconds, flagging one potentially fake citation and several assertions unsupported or contradicted by the cited cases, with no signup and no data stored.