Track 2, Mihir Milind Ughade, SentinelDesk

[mihir0804]

Hey everyone, here is my submission for Track 2: SentinelDesk.

I built SentinelDesk because I wanted a truly local, autonomous desktop AI that actually takes full advantage of AMD hardware without relying on cloud APIs. It's designed to act as a hands-on system copilot—meaning it runs entirely on-device, keeps user data completely private, and can execute system-level commands on its own.

Here is a quick breakdown of how it works under the hood:

Native ROCm Inference: The LLM engine is built explicitly on PyTorch customized for AMD ROCm. There are no external server calls, ensuring zero latency and maximum hardware utilization.

Resource Friendly: To make sure it runs smoothly on consumer Radeon GPUs without hitting VRAM bottlenecks, I integrated bitsandbytes to load models (like Llama-3-8B-Instruct) in 4-bit or 8-bit precision.

Persistent Memory: It's not just stateless text generation. I hooked up a lightweight ChromaDB instance so the agent remembers past interactions and user persona settings across different sessions.

Autonomous Tool Execution: The copilot goes beyond chatting by actually running local system diagnostics and file management tasks autonomously when needed.

I've detailed the full architecture and specific AMD optimization strategies in the Project_Specification.md and README.md files.

Demo Video: https://youtu.be/ti3ntYJ2NO4

Looking forward to sharing this with the judges and the rest of the community!