Private
Pilot
Pilot is a natural-language agent that controls real applications on your behalf — across the web, Windows desktop, and Android — through one interface that never needs to be taught a new app.
PythonPlaywrightGemini APIfaster-whisper
One executor, three surfaces
Rather than integrating with each application individually, Pilot exposes a single device-agnostic executor interface backed by three drivers: the DOM for web, Windows UI Automation for desktop apps, and ADB for Android. Whatever the task touches, the same planner can operate it.
Reads structure, not pixels
Instead of taking screenshots and guessing at coordinates, Pilot reads accessibility and DOM trees, cutting per-step token cost roughly 2–5x and letting the planner target semantic element IDs instead of brittle pixel positions — with a visual Set-of-Marks fallback for canvas-based UIs that don't expose a tree.
Observe, plan, act, verify
Pilot batches 3–5 actions per model call, verifies each one against a tree diff, and self-recovers through retry and replanning with stall detection — so a single wrong click doesn't derail the whole task.
Offline voice and safety controls
A fully offline voice pipeline (VAD, local speech recognition, wake word, text-to-speech) runs at roughly 430ms speech-to-text with no audio leaving the machine. Irreversible actions require confirmation, credentials are handed off rather than typed by the agent, a system-wide kill switch is always available, and every action is written to a full audit trail.
Interested in working together?
laksh@lakshjain.com