After my first post, I wanted to ask the AMD AI community for some advice.
Over the last four months I’ve been building Nova, what I now describe as an Autonomous Reasoning Operating System.
Nova wasn’t built on an existing AI framework or agent SDK. She grew from a different goal: creating a long-lived local system that can reason about work, govern its own execution, measure evidence, and improve its future decisions while keeping the operator in control.
Today Nova already:
-
Selects models based on the hardware she’s running on.
-
Understands GPU, VRAM, and NPU availability through her SOCK (System Optimization and Compatibility Check) subsystem.
-
Tracks governed work instead of conversations.
-
Measures progress using evidence instead of assumptions.
-
Writes a growing portion of her own implementation under governance while I focus on architecture and mission design.
Now I’m at the point where I want to start taking advantage of AMD’s AI ecosystem instead of treating Ryzen AI as just another processor.
This is where I’d really appreciate guidance from people with experience building on AMD hardware.
I’m particularly interested in:
-
Integrating the Ryzen AI NPU into Nova’s reasoning pipeline.
-
Understanding where AMD’s AI software stack fits into a long-running autonomous runtime.
-
Learning how others are balancing CPU, GPU, and NPU workloads for local AI systems.
-
Any tools, SDKs, profiling techniques, or best practices that would help move Nova forward.
The repository is public because I believe architecture should be inspectable, not just demonstrated.
I didn’t build Nova to prove a point. I built her because I wanted my own Jarvis. Four months later I’m building something much larger than I ever expected, and I know there are people in this community with knowledge that can help shape where Nova goes next.
If you’ve worked with Ryzen AI, XDNA, ROCm, or AMD’s AI tooling, I’d genuinely appreciate your thoughts.