# VOID — AI Screen Assistant > A floating AI desktop assistant that reads your screen, understands Telugu, and helps you reply, summarize, translate, and draft — powered entirely by local models.

---

Overview

VOID runs as a frameless floating widget on your Windows desktop. It uses Qwen2.5-3B (GGUF, CPU-only) for all text tasks and Groq Vision for screen understanding. No cloud LLM dependency for text generation — everything runs on your machine.

---

Features

Action	Description
LangGraph Agent	Multi-step task planning and execution
WhatsApp Reply	Captures chat, generates 3 Tenglish suggestions via Ollama Vision
Screenshot	Saves PNG to ~/Pictures/VOID/
Voice Type	Transcribes mic via Whisper, types into active window
Explain Screen	Analyzes screen with local vision model (llava/moondream)
Translate	Extracts and translates on-screen text
Summarize	Summarizes screen content
Email Draft	Voice note to professional email
Ask VOID	Persistent chat with memory across sessions
Alt+Space Hotkey	Summon VOID from anywhere

---

Architecture

void/
├── project/
│   ├── backend/
│   │   ├── main.py                  # FastAPI + LangGraph agent
│   │   ├── config.py               # Environment configuration
│   │   ├── agent/
│   │   │   └── void_agent.py       # LangGraph agentic core
│   │   └── services/
│   │       ├── ollama_service.py   # Ollama LLM client
│   │       ├── vision_service.py   # Local vision (llava/moondream)
│   │       └── memory_service.py   # SQLite RAG memory
│   └── frontend/
│       └── void_ball.py            # PyQt6 glassmorphism UI

---

Tech Stack

Layer	Technology
Agent	LangGraph 0.0.x — multi-step planning
LLM	Qwen2.5 via Ollama (local)
Vision	Ollama llava or moondream (local)
Memory	SQLite RAG — persistent context
ASR	OpenAI Whisper tiny (local)
TTS	pyttsx3
Backend	FastAPI + Pydantic
Frontend	PyQt6 with glassmorphism

---

Setup

1. Prerequisites

# Install Ollama
winget install Ollama.Ollama

# Pull models
ollama pull qwen2.5:3b
ollama pull llava

2. Install dependencies

pip install -r requirements.txt

3. Start Ollama (in separate terminal)

ollama serve

4. Start backend

cd project/backend
uvicorn main:app --host 0.0.0.0 --port 8000 --reload

5. Start frontend

cd project/frontend
python void_ball.py

---

API Endpoints

Method	Route	Description
POST	/agent/query	LangGraph agent (multi-step planning)
POST	/agent/simple	Direct LLM query
POST	/vision/analyze	Analyze screenshot
POST	/vision/explain	Explain screen content
POST	/vision/whatsapp-suggest	Generate WhatsApp replies
POST	/vision/save-screenshot	Save screenshot
GET	/memory/history	Conversation history
GET	/memory/actions	Action history
POST	/memory/remember	Store important fact
POST	/memory/recall	Retrieve relevant memories
GET	/health	Service status

---

Configuration

Create project/backend/.env:

OLLAMA_HOST=http://localhost:11434
OLLAMA_MODEL=qwen2.5:3b
VISION_MODEL=llava
SCREENSHOTS_ROOT=C:\Users\You\Pictures\VOID

---

Dependencies

• LangGraph
• Ollama
• Qwen2.5
• llava
• PyQt6
• OpenAI Whisper

---

VOID — local-first AI assistant for Telugu users.