I work on security architecture and agentic identity, and I use my home lab to test AI systems firsthand. This page is a running record of what I’ve built, what I’m testing, and what I’ve learned. I’ll update it as the setup changes.
What’s in the lab
My main local AI machine is a MINISFORUM AI X1 Pro with an AMD Ryzen AI 9 HX 370, Radeon 890M integrated graphics, and 96 GB of memory. I run local models through Ollama. The wider environment includes a Linux server, a Synology NAS, Home Assistant, and n8n.
I also use Hermes and LifeOS as I explore how an assistant can work across my notes, tools, and everyday workflows. Paperclip is a more recent experiment in coordinating agents. These components are at different stages of maturity; being in the lab does not mean I’ve put them all into production.
Cloud models and local routing
I use both local and cloud-hosted models. Ollama runs models on my MINISFORUM, while Amazon Bedrock gives me access to remote models. I run LiteLLM as a local proxy between my AI tools and cloud model providers. This gives the lab a place to configure model access without putting provider-specific connections into every tool.
I’m still documenting which workloads work best locally and which benefit from a cloud model. My choice depends on the task, model capability, latency, cost, and where the data will be processed.
Architecture at a glance
How it developed
2024 — Workflow ideas. I began exploring ways to monitor YouTube content, extract transcripts, analyze them with fabric, and return useful results through tools such as Slack. These were design discussions, rather than a confirmed deployed system.
Spring 2025 — A dedicated local AI machine. The MINISFORUM AI X1 Pro became part of the lab. I ran Ollama on Windows and experimented with connecting workloads in Windows Subsystem for Linux (WSL) to its models. This exposed practical issues around GPU detection, model compatibility, timeouts, and tool-call formatting.
Fall 2025 — Automation and local models. By June, I had n8n installed at home and connected to Home Assistant. During the year, I worked with Ollama, Open WebUI, Docker, and model-serving options. I also designed multi-agent demonstrations using Model Context Protocol (MCP), transcript analysis, Linux commands, and document creation. Some were prototypes or plans; I’m keeping those distinct from the services I know were running.
September 2026 — Getting GPU inference working. My first Ollama runs with Qwen3-30B-A3B were largely CPU-bound. After investigating the AMD driver and integrated GPU configuration, Ollama reported all 49 model layers offloaded to the Radeon 890M. It was a useful reminder that a model loading successfully does not prove it is using the hardware I intended.
September 2026 — Connecting the pieces. I’ve been working with LifeOS on my laptop and Linux server, thinking through how to sync its memory and use it with Hermes. I’ve also explored MCP integrations for practical tasks such as publishing to wordpress. More recently, I started configuring Paperclip for network access and comparing agent workspaces and orchestration tools.
What I’m learning
Measure the actual workload. Model size, available memory, GPU offload, context length, and response speed all matter. A successful launch is only the start of a useful test.
Integration is harder than a demo. Getting an assistant to call a tool once is different from giving it dependable access, durable memory, clear permissions, and recoverable failures.
Security belongs in the architecture. My professional work in identity and security shapes the lab. I’m especially interested in what authority an agent has, whose identity it acts under, what data it can reach, and how its actions can be audited.
Keep an honest changelog. I want this page to show working configurations, failed experiments, and decisions I reconsidered. I’ll add dated updates as I test new models, tools, and integrations.
Lab notes and updates
September 2026: Established GPU offload for Qwen3-30B-A3B on the Radeon 890M; continued work on Hermes, LifeOS, and Paperclip. My current lab also connects to Amazon Bedrock and other remote models through a local LiteLLM proxy; the deployment dates of those components are not recorded here.
Last reviewed: September 27, 2026.