14 open-source projects use Jev for agent decisions on GitHub
A roundup of 14 projects including jev-trader, jev-ultrafast and jev-doom-agent that use a typed model instead of an LLM for if-statements, cutting latency from 2300ms to 180ms.
this is pure f*cking fantastic Jev has been exploding across GitHub since launch: 14 open-source projects just proved your agent doesn't need a 2,000-token conversational LLM to evaluate an if-statement. Over 80% of agent latency is burned on decisions requiring zero natural language: "which tool next?", "is this diff risky?", "should I execute?". Outsourcing deterministic logic to heavy frontier models creates a 2,300ms latency penalty and a 10x API bill for decisions a typed model makes in 180ms for $0.0001. None of these projects generate a single word of text. Every one of them returns a deterministic probability tensor against an answer defined in code. high-frequency trading. browser clicking. wasm doom. dinosaur reflexes. typesafe chess. drone navigation. tax form classification. Here is Part 1 of what developers are already shipping on GitHub with Jev: 01 jev-trader — real autonomous trading bot placing live limit orders on Monad every 300ms block, judged by Jev alone with zero text model latency. 1,911 stars. https://github.com/jarrodwatts/jev-trader 02 jev-ultrafast — browser automation agent that picks every DOM click itself in 180ms, only waking up a heavy language model when it actually needs to type something. 16,758 stars. https://github.com/browser-use/jev-ultrafast 03 jev-doom-agent — real Chocolate Doom compiled to WebAssembly. Two engines running the same map in parallel, with Jev picking tactical combat macros every single frame. https://github.com/lukaske/jev-doom-agent 04 jev-t-rex-runner — the Chrome dinosaur game played with sub-50ms reflexes. Jev continuously evaluates obstacle proximity to pick jump, duck, or run. https://github.com/joshlarsen/jev-t-rex-runner 05 typesafe-chess — Jev vs a search engine, colors swapped. The search won, but overruled Jev's first instinct on roughly half the moves. https://github.com/TholeG/typesafe-chess 06 jev-drone — simulated quadrotor clears a five-station obstacle course by camera alone, with Jev making spatial navigation verdicts twice a second. https://github.com/RomanSlack/jev-drone 07 tax-doc-classifier — sorts real IRS tax forms with 100% strict accuracy across 261 forms at roughly $0.001 a page, eliminating OCR prompt hallucinations. https://github.com/kyotofin/tax-doc-classifier the loop: raw event bus → candidate list assembled in code → parallel typed tensor (180ms p50) → calibrated confidence gate (>0.85) → deterministic execution the telemetry: 2,300ms p50 latency collapsed to 180ms. $0.03 per if-statement collapsed to $0.0001. A conversational chatbot writes essays. A typed decision runtime executes. Part 2 (MCP Servers, Codex Routers, Context Garbage Collection & Postgres) below ↓
This post quotes another post: x.com/damkina7/status/2102860956724335039