Overview
I trained and deployed a small, specialized local decision model engineered specifically for macOS desktop automation workflows. Rather than relying on large cloud-hosted models for OS-level interactions, this project explores running fast, private, and deterministic agentic decision-making locally on consumer hardware.
I built the complete machine learning pipeline independently from the ground up. To maximize representation efficiency for system operations, I trained a custom domain-specific tokenizer tailored to macOS file paths, shell syntax, keyboard shortcuts, and UI automation semantics.
With the specialized tokenizer, I pre-trained a compact transformer architecture on structured execution data, followed by targeted fine-tuning on sequential decision-making and multi-step desktop automation tasks. To ensure smooth local performance, the model was quantized and compiled for Apple Silicon, leveraging unified memory and hardware acceleration to deliver low-latency on-device inference with minimal energy overhead.
Demonstration
The video below shows the local decision model executing automated decision tasks directly on macOS in real time.