The goal
Build a self-hosted assistant on hardware I control, while learning to manage the full path from firmware, storage, and power management to guest operating systems, accelerators, model serving, and user interfaces.
The architecture
A Ryzen-based workstation runs Proxmox VE. An Ubuntu Server VM receives an NVIDIA RTX 3060 through PCIe passthrough to support local Ollama inference. User-level services are used for a developing voice interface and other integrations.
Real troubleshooting
I worked through boot instability, firmware settings, GPU visibility, driver installation, guest/host service availability, API authentication, and resource/uptime observations. A process reporting “active” is not the same as a responsive application, so endpoint verification became part of testing.
Where the project is going
Structured system monitoring, personalized briefings, more reliable voice interaction, durable knowledge retrieval, and optional integration with Watchtower and real-world weather/traffic feeds.
Technical boundaries
This is a self-managed lab rather than a production enterprise MLOps installation. I use AI-assisted programming tools, but I am responsible for design constraints, integration, functional testing, and investigating failures.