Daily workstation
RTX 5090
Blackwell · 32 GB VRAM
Primary workstation GPU for daily inference, image and video generation, and testing local AI workflows.
AI Infrastructure
Local compute.
Connected intelligence.
My daily AI stack spans an RTX 5090, RTX 4090, and a two-node DGX Spark cluster. Local model serving, image and video generation, and Hermes orchestration, with cloud models available for the work that benefits from them.
01 / Compute
The hardware behind the work.
Three primary systems for everyday inference, creative workloads, and cluster experimentation.
The workstation, dedicated GPU VM, and Spark cluster give me different places to run and experiment with workloads. I treat them as distinct compute resources, with model size and the task guiding placement. That keeps the hardware overview tied to how I use the systems, from a local generation workflow to serving a model for an agent.
Dedicated GPU workloads
RTX 4090
24 GB VRAM · Proxmox
A second daily-use GPU, passed through to a Proxmox VM for inference and GPU workloads alongside the workstation.
Cluster compute
2× DGX Spark
GB10 · 200 Gb/s interconnect
Two GB10 systems with 128 GB unified memory each. The primary serves the vLLM API and coordinates distributed inference with the worker.
The Sparks communicate over a dedicated 200 Gb/s ConnectX-7 link for distributed inference. Their 256 GB of aggregate unified memory remains split between two nodes, with the serving configuration distributing the work. See how the cluster connects →
Additional lab capacity includes an RTX 3080, RTX 3070, and two RTX 2060s for smaller workloads and testing. See the wider lab →
02 / Model Serving
Local models. Useful endpoints.
vLLM and SGLang are the focus of my serving stack, making local models available to the tools and agents I use.
vLLM
Local LLM inference and API serving, including workloads on the DGX Spark cluster.
SGLang
An additional serving engine in my daily stack for running and evaluating local language models.
CUDA
The GPU compute foundation for inference and generative workloads across the NVIDIA systems.
Model choice and placement depend on memory requirements, workload, and the capabilities each client needs. Keeping both vLLM and SGLang in the stack gives me room to evaluate the serving layer alongside the model. For an agent, that includes how the client interacts with the endpoint; for interactive work, it includes how usable the response feels. Local endpoints and cloud providers both have a place in the setup.
03 / The Request Path
One gateway. The right backend.
Hermes, Open WebUI, and development clients send model requests through LiteLLM. The gateway routes the request to a configured inference endpoint, including the primary node of the Spark pair.
LiteLLM routes the request
Named model routes give clients a consistent entry point while keeping serving separate from the interface. GPU execution stays on the backend; the gateway handles model access and routing.
Hermes manages the work
Hermes combines the instructions, memory, and tools around each model call. A response can inform the next step, while configured permissions and approvals control tool execution. Results become context for the next iteration.
04 / Image & Video
From a prompt to a visual workflow.
Image and video generation are part of my daily GPU use, alongside language models and coding agents.
Creative tooling
ComfyUI
I use ComfyUI for image and video generation, working with node-based workflows to connect models and generation steps, adjust settings, and iterate on outputs.
The workflow itself is part of the work: it records how the generation steps fit together and gives me a starting point for the next iteration. That makes it easier to explore a visual direction, revisit the setup, and adjust individual stages as I work toward an output.
Image generation Video generation Workflow iteration
05 / Agent Orchestration
Hermes connects models to tools.
The agent layer brings instructions, memory, and tool use together. Hermes is central to how I orchestrate that work across local and cloud models.
Hermes Agent
The orchestration layer for agent work, bringing together models, tools, memory, and reusable skills.
OpenShell
Policy-controlled execution environments for agent tools and workloads.
OpenClaw
An additional assistant and agent environment for coding and automation tasks.
Reusable skills capture procedures that are worth repeating. I think about the orchestration layer in terms of the instructions an agent receives, the tools it can call, and the context it needs to continue useful work. The model is one part of that setup. Model selection can change with the task, while the surrounding tools, skills, and workflows remain the focus.
06 / Development
AI in the development loop.
Coding tools I use to implement changes, investigate problems, and review work across my projects.
Claude Code
Daily development work across repositories, from implementation to debugging and review.
Codex
Coding and review alongside Claude Code, including second opinions on implementation decisions.
OpenCode
A coding interface for working with local model endpoints and hosted providers.
07 / Workflow Automation
Connect the steps. Repeat the useful work.
n8n connects models and services to the events and schedules that make a workflow useful beyond a single session.
n8n
I use n8n to connect APIs, call models, transform results, and pass them to the next service. It provides the workflow layer alongside Hermes agent orchestration.
The focus is on repeatable integrations and background jobs, with local or cloud inference selected for each workflow.
Put it to work
Build your AI workflow.
From local model serving to image generation and connected automation, I can help design the infrastructure and integrations around your work.