Applications

jev-canvas

@gaborishka★ 2JavaScriptMITupdated 2026-09-19▶ runnable

Draw on a tldraw canvas with your voice and a pointing finger. Jev (TypeSafe System One) decides action, target and place in ~350 ms per spoken word.

gaborishka/jev-canvas ↗

Where it calls Jev

<Card title="JEV" note={view.busy ? 'asking…' : jev ? `${jev.latencyMs} ms` : ''} tone={jev?.verdict}>

src/App.jsx:170

The link points at the commit we read, so the line number still holds.

What it asks Jev

These question sets are lifted from this project's source exactly as written.

from shared/questions.js:125

  1. is_commandtrue/false

    Is the speaker giving an instruction to the drawing canvas? The speaker may use English or Ukrainian.

  2. completetrue/false

    Has the speaker finished the instruction, so that it can be carried out right now?

Run it on your own text

The code
import { experimental_evaluate as evaluate } from 'ai';

const { answers } = await evaluate({
  model: 'typesafe-ai/jev',
  state,
  questions: {
    is_command: { type: 'boolean', instructions: 'Is the speaker giving an instruction to the drawing canvas? The speaker may use English or Ukrainian.' },
    complete: { type: 'boolean', instructions: 'Has the speaker finished the instruction, so that it can be carried out right now?' },
  },
});

Other projects in this category