Jev Driver
Browser driving game where Florence-2 captions dropped images and Jev decides whether the car stops, slows or continues.
Overview
Jev Driver asks whether Jev can drive a car. In a top-down browser game you drag images of pedestrians, obstacles or traffic signs onto the road or sidewalk. Florence-2 runs in the browser through Transformers.js and writes a short caption; only the caption and the object's location go to a Cloudflare Worker, which asks Jev to choose an action such as stop, slow_down, wait_for_green or change_speed. The game handles movement.
A list summary adds that the worker asks three Choice questions (action, category and speed limit) and that no rule table overrides the answer. The README reports a tuning run with about 330 ms median latency, the expected call in 151 of 156 scored scenarios, and roughly $0.00004 per decision. A Thoughts panel shows Jev's confidence, and the author notes it still makes mistakes. The live demo needs WebGPU and downloads about 357 MB of model weights on first visit.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | What should the car do about this object? | stop · slow_down · wait_for_green · change_speed |
| Choice | What category is the object, and what speed limit applies? | — |
How it works
- User drops an image into a lane or sidewalk
- Florence-2 captions it in the browser
- Worker sends caption and location to Jev
- Game applies the chosen action and shows confidence
Strengths
- Images never leave the browser
- Low per-decision cost reported
- Live public demo
Caveats
- Tuning-run results only
- Large first download; CPU fallback is slow
- Game, not a real driving system
Getting started
Open https://drive.mrza.ch in a WebGPU browser and drag images onto the road.At a glance
- Type
- Project
- Status
- prototype
- Language
- TypeScript
- Runtime
- Browser with Cloudflare Worker
- Has code
- yes
Use cases and integrations
How curated lists describe it
Simulation: a top-down driving game where Florence-2 captions each image dropped on the road in the browser and a Cloudflare Worker asks Jev three `Choice` questions (action, category, speed limit) about the caption and its lane or sidewalk, with no rule table overriding the answer; live at drive.mrza.ch, about 330 ms median and $0.00004 per decision.
yibie/awesome-jev
Frequently asked questions
What does Jev decide?
How the car should respond to each captioned object, plus its category and speed limit.
How accurate was it?
The README reports the expected call in 151 of 156 scored scenarios in a tuning run.
Are my images uploaded?
No, only the caption and location are sent.
Related projects
Real-time bot battles where you write a fighter's strategy in plain English and a System One model pilots it several times a second.
ProjectbetaTypeScriptBot that plays Red Alert 2 web skirmishes, with code handling mechanics and a decision model making stance, threat and timing calls.
ServicebetaPythonMultiplayer 3D science-fantasy version of Houston's River Oaks District with Jev-powered NPC companion, chauffeur and smart drive.
ProjectbetaTypeScriptBrowser 3D planetary rover sandbox where a TypeSafe controller chooses the next action from complete candidates, versus a baseline.
ProjectbetaTypeScript- clash-jev★ 36
Clash Royale bot on a real phone with no trained policy: vision builds a JSON state and Jev picks strategy, card and square each second.
ProjectprototypePython Doom runs as a spacecraft payload over real flight software; Jev pilots from the ground every half second while Claude reviews.
ProjectprototypePython