Sahayak

README excerpt

# सहायक · Sahayak

**An on-device Android agent that operates your phone for you — and knows when to stop and ask.**

Built for the Google DeepMind Bangalore Hackathon (Gemma 4 track), 11 July 2026.
Everything runs on the phone: **Gemma 4 E4B via LiteRT-LM**, on-device STT/TTS, a local
SQLite memory. There is no server, and the app makes no network calls.

Say *"आर्यन को मैसेज भेजो कि मैं कल घर आ रहा हूँ"* and it opens Messages, finds the
contact, types the message — then **stops at the Send button and asks you**.

---

## Why this is an agent, not a chatbot with a microphone

The Gemma track asks for the full **sense → decide → act → check** loop, local error
recovery, local state, and clear boundaries for deferring to a human. That's the whole
design:

```
SENSE     accessibility tree → prune to interactive nodes → number them (Set-of-Mark)
  ↓
REPLAY?   a learned step for this screen? → execute it, skip the model entirely
  ↓
DECIDE    Gemma picks ONE action from the numbered nodes (constrained JSON)
  ↓
GATE      irreversible? (send / pay / delete) → stop, ask the human, wait
  ↓
ACT       tap / long-press / type (SET_TEXT, replace-semantics) / scroll / open-app
  ↓
CHECK     wait for the window to settle → re-sense → did the subgoal actually happen?
  ↓
journal   persist the step, advance — or recover
```

The model chooses one action per screen. It never plans blind: subgoals come from a
plan made once per task, and every step is verified against the real screen afterwa