The Desk
The workstation this site, its code, and most experiments run from. Two environments — a rack of dedicated servers for the cluster, and a local NVIDIA GB10 Blackwell rig for ML workloads that shouldn't leave the room.
Cluster
Provisioned with Ansible. Everything reproducible from a repo — no artisanal servers.
- Control plane
k8s cluster
Dedicated Hetzner servers. All applications live in the `applications` namespace. Ingress-nginx at the edge, cert-manager for TLS.
- Messaging
Self-hosted Matrix
Synapse + Element web. Federated chat without Telegram/Slack lock-in. Also used as the notification bus for internal webhooks.
- Automation
n8n
Pipelines for webhook fan-out, scheduled scraping, and stitching the ML services together (transcript → summarize → notify).
GB10 — under the desk
NVIDIA GB10 Blackwell. Enough for local inference on mid-size open-source models without sending data out.
- Speech-to-text
Whisper
Real-time transcription of meetings and voice notes. Feeds the automation layer.
- Image generation
ComfyUI
Node-based workflows for experiments, thumbnails, and visual prototyping. Custom checkpoints and LoRAs.
- Local LLM
Ollama
Qwen and Gemma 3 as the local models. Chat, code completion, and agentic workflows without sending prompts over the internet.
Why
Three reasons. First — keep prompts and voice off third-party APIs. Second — running the stack locally forces you to understand it, which pays back on production systems at work. Third — it's fun.