Moritz Kremb built a voice-controlled browser. Jev selects the intent and target from partial transcripts, while Playwright performs the browser action.
Input
Speech transcript and a compact snapshot of page elements.
Jev
Choose the command intent and a target element.
App
Application policy checks the choice; Playwright then acts or waits.
Third Hand reads available controls and local screen text, asks Jev which action fits a request, and carries out the selection through macOS automation.
Input
A user request and descriptions of accessible on-screen controls.
Jev
Select a matching action from the available controls.
App
The macOS app clicks, types or checks the resulting state.
Rolewise retrieves job descriptions separately, then asks Jev whether each resume and job pair is suitable. The author shared a recording of the requests.
Input
A resume and a retrieved job description.
Jev
Select suitable or not suitable for each pair.
App
The app narrows the opportunity list for human review.
The inspected X post includes a video demo but did not link public code for this workflow.
The clip does not establish how the workflow behaves for ambiguous files or conflicting rules.
Worth testing: What safeguards prevent a wrong decision from moving or renaming an important file?
Navigation / SimulationLive demo
Choosing navigation actions for a lunar mission
By Rahman Yoonus
Odyssey is a browser simulation of an Earth-to-Moon journey. The simulation provides trajectory and maneuver timing; Jev chooses bounded attitude and correction actions at each step.
Input
Spacecraft state: current trajectory, maneuver timing, and available navigation actions.
Jev
Jev selects an attitude or correction action from the bounded action set.
App
The simulation advances the spacecraft state and loops to the next decision.
A deterministic fallback activates when the model is unavailable. The creator notes this without specifying which actions the fallback takes.
WebGL2 is required to run the demo; reported action quality is limited to what the live overlay shows.
No source code or repository is linked; only the live demo is available to inspect.
Worth testing: How does the action selection change when the available action set is reduced? Comparing full vs. limited sets would show how much the selection relies on the choice structure.
Axis definitions are unconstrained; a vague or contradictory axis produces a meaningless map.
The score reflects Jev's interpretation of the axis phrases, not an objective measure of semantic distance.
No source code or repository is linked; only the live demo is available to inspect.
Worth testing: Do different axis definitions produce consistent relative positions for the same set of phrases? Try reversing an axis phrase and check whether the positions invert as expected.
Official and integration examples
Provider demonstrations and integration walkthroughs. Not independent replications.
