Does Jev actually see the road images?

Jev never sees pixels. Florence-2 captions each image locally and only the caption plus location go to a Cloudflare Worker that calls Jev. Caption errors therefore carry straight through to the driving call.

How reliable is the driving behaviour in the tests?

The tuning run matched the reference on 151 of 156 scored cases with 330 ms median Jev time, but four prompt-injection probes had no reference and were excluded. Misses like the teddy bear and amber light are documented as part of the experiment.

What setup catches should I expect before playing with the demo?

Use a WebGPU browser, wait for the model download to cache, and try objects in different lanes or sidewalks. A slower CPU fallback exists but captioning may lag the car.

What leaves my machine, and what licence covers the pieces?

Sample photos retain their Wikimedia Commons licences under the credits file, while the code itself is MIT. Image data stays in the browser; only captions and locations are sent server-side.

What do I need to run it locally?

Node 22.18+ on the Node 22 release line, pnpm 11, and a Cloudflare account with Workers AI and prepaid AI Gateway credits, since the dev server calls real Jev; without credits calls fail with 2021 Insufficient AI Gateway credits.

How heavy is the vision model?

The first visit downloads about 357 MB of weights, cached afterwards, and Florence-2 captions at 384 px in about 0.4 seconds on an M1 with WebGPU. A 228 MB CPU fallback exists but captioning is much slower.

What happens when a call fails mid-drive?

If captioning or the decision request fails, the game stops for objects in your lane, slows for the oncoming lane and continues past the sidewalks.

CommunityJev Demos & Experiments

Vision Jev

Browser driving game where an in-browser vision model captions dragged-in road objects and Jev maps each caption plus location to stop, slow down, or continue.

See highlights

Open github.com

Stars: 4
Language: TypeScript
License: MIT
Published by: Community
Author: reinhard-z
Last verified: Oct 2026

reinhard-z/vision-jev

User repository on GitHub

View on GitHub

Browser driving game that captions dropped road images locally, then uses Jev Choice answers to drive the car while code owns physics and timing.

Forks: 0
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

Can Jev drive a car? An experiment with Jev, a fast classification model from TypeSafe, to see what it can do when decisions have to happen in real time. I built a browser game where it drives a car. You drag images of pedestrians, obstacles or traffic signs onto the road. A vision model in your browser writes a short caption for each image, and Jev uses that caption and the object's location to decide whether to stop, slow down or keep going.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import

Works forJev for Classification

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Chains browser Florence-2 captions plus object location into Jev driving-action decisions.
Game movement, braking, and timers run in code with no rule table overriding Jev answers.
Documents tuning cases, timing splits, failure examples, and vision-model comparison data.
Runs locally with Node, pnpm, and a Worker, with game, pipeline, and Worker test commands.

Watch out

First visit downloads a large vision model that is cached afterwards, and local runs need Cloudflare Workers AI with prepaid gateway credits. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

laya-server

Self-hosted Docker service and web UI that packages Laya structured-decision inference behind a compatible System One endpoint.