Tech
I sit somewhere between a clinician who codes and a sysadmin who reads brain MRIs. These are notes on the tools and systems I use in clinical work, research, and infrastructure.
AI in Clinical Radiology
At Turku University Hospital, we have been integrating commercial AI algorithms into emergency radiology workflows since 2022. In 2025 we completed a 6-month HTA for AI algorithms doing fracture and effusion detection. For this project I also built "Pekka 2000", a QC system for radiologists to collect reports about AI results and to produce live statistics. The platform continues to collect metrics for AI systems for QC. We are now expanding our use of commercial AI systems that prove useful in clinical practice. We use AI agents to build agent-driven workflows for AI research and to build small apps for our own use.
Sauhu
Sauhu is a DICOM viewer I built in Rust for my own clinical use. Multi-viewport with synchronized scrolling and reference lines, MPR reformatting, GPU-accelerated volume coregistration, measurements, full PACS connectivity, and CT presets. Designed for Omarchy on Wayland. Everything a radiologist needs in one fast app.
Infrastructure
- Daily driver: Framework Laptop 13 (Intel Core Ultra 7, 96 GB RAM) running Omarchy, an opinionated Arch Linux distribution with Hyprland on Wayland.
- Home server: Intel i7-12700K, 64 GB RAM, running Omarchy. Hosts multiple VMs using KVM, including the one serving this website. Caddy for web, Soft Serve for private Git repos.
- Hospital servers (Ahjo and upcoming Masuuni): i9-14900K with RTX 5090 (two RTX 5090 cards in Masuuni), running Podman containers and KVM VMs. VMs host commercial radiology AI software in production (Radiobotics, RapidAI, etc.). Also used for local LLM inference with Ollama, DICOM processing, and AI model development. Twin server setup for HA, built with AI agents communicating across both servers. Basic sysadmin work is done with local and cloud-powered AI agents, using PHI-safe tools built just for agents.
Why self-host? Partly for privacy, control, and costs, partly because managing Linux servers is how I learned most of what I know about networking, security, and systems dealing with DICOM and HL7. I have maintained dcm4chee setups, also using Pacemaker and DRBD for HA, for over ten years. When something breaks at 2 AM, you are suddenly a fast learner. Nowadays AI agents help a lot.
Tools
- Editor: Neovim
- Shell: Bash + Starship prompt
- Terminal: Ghostty + tmux
- AI assistant: Claude Code for sysadmin tasks, coding, and research
- External brain: Obsidian (GTD workflow)
- Presentations: Marp (Markdown to slides), exported to PowerPoint or PDF
- Medical imaging: Sauhu (DICOM viewer), pydicom, pynetdicom for DICOM scripting
- Version control: Git + GitHub CLI, self-hosted Soft Serve for private repos
- Database: PostgreSQL (production), SQLite (prototyping)
DICOM and Hospital IT
Medical imaging runs on DICOM, a protocol from the 1990s that somehow still works. Most of my technical work at the hospital involves connecting things to the PACS: routing studies to AI analysis VMs, building audit logging for image access, and automating data flows between systems that were never designed to talk to each other.
The challenge is always the same: hospital networks are locked down for good reason, vendor systems are black boxes, and every integration is a negotiation between security, usability, and the reality of a 24/7 clinical environment.