Back to Portfolio
Sole Developer
Ongoing Research
aiOS (Privacy-First)
Privacy-First Computer Use Agent for On-Device Automation
Demo Video Placeholder
The Challenge
Existing CUAs rely on cloud models, creating massive privacy risks and latency issues. Vision-only approaches are slow and error-prone.
The Solution
aiOS hybridizes Accessibility Tree (AX) parsing with localized VLM verification. It runs small, specialized models (Function Gemma, Qwen) locally on Apple Silicon via MLX.
System Architecture
Architecture Diagram: placeholder_aios_arch.png
Tech Stack
Python
Apple MLX
Neo4j (GraphRAG)
Qwen 2.5
Function Gemma
Key Features
- 100x Faster than Vision-only agents via AX Tree
- Zero data egress (Local Inference)
- Self-correcting workflow using GraphRAG memory