Engineering software from application code to production infrastructure.

swisslex.mdProfessionalSwisslex AG

Fullstack and platform engineering

At Swisslex I own features of the Swisslex platform end to end: business logic and APIs in C#/.NET, user interfaces in Angular and TypeScript, deployment and operations. Architecture discussions and code reviews are part of the daily work.

I also work on the platform those applications run on: Kubernetes with GitOps-based delivery through Helm and Argo CD, CI/CD automation in GitLab, an observability stack built on Prometheus, Grafana and Loki, and troubleshooting in distributed environments. I also drive technology evaluations for the platform, weighing options against our operational requirements and turning the results into recommendations the team can act on.

Coding agents are part of my workflow for refactoring, infrastructure work and debugging across repositories. Technical judgement, code review and hands-on debugging stay with me.

• C# / .NET
• Angular
• TypeScript
• SQL Server
• Kubernetes
• Helm
• Argo CD
• GitLab CI
• Prometheus
• Grafana
• Loki

// Generic illustration. Internal systems are not described here.

wsl-research.mdProfessional · part-timeWSL

Research software for quantitative wood anatomy

At the Swiss Federal Institute for Forest, Snow and Landscape Research WSL I develop ROXAS AI, an open-source napari plugin for quantitative wood anatomy. It detects tree rings and wood cells in microscopy images with deep learning, measures traits such as lumen area, cell wall thickness and ring width, and turns them into research-ready datasets for climate and forest research.

My work covers the segmentation of cells and ring boundaries, interactive manual editing of the results, crossdating, measurements, GPU support across CUDA, Apple Silicon and Intel GPUs, and the user documentation. The plugin is in active development and not yet officially released.

• Python
• napari
• PyTorch
• CUDA · MPS · XPU
• NumPy
• pytest

• Repository github.com/roxas-ai ↗
• Documentation roxas-ai.github.io ↗

local-agents.mdIndependent · CAS projectpidev

Local AI and sandboxed coding agents

pidev runs a coding agent inside an NVIDIA OpenShell sandbox on a Mac, backed by a model served with vLLM on an NVIDIA DGX Spark. One command per project, and the agent only ever sees that project folder.

Landlock restricts the file system, a per-binary network policy controls outbound traffic, and Git tokens never enter the sandbox. Custom TypeScript tools extend the agent; web access runs through a self-hosted SearXNG and fetch service with an SSRF guard, and a persistent memory service combines semantic (bge-m3) and full-text search.

• DGX Spark
• vLLM
• Qwen
• OpenShell
• Landlock
• TypeScript
• Python
• Docker

local-rag.mdIndependentGoddoniWiki

Retrieval-augmented generation on local hardware

GoddoniWiki answers questions against my own Obsidian notes without any cloud service: bge-m3 embeddings, a ChromaDB index, Qwen3-14B on vLLM and an Angular front end that streams the answer and shows its sources.

A second setup exposes the same kind of index to an agent through an MCP server, which let me compare a fixed retrieval pipeline with retrieval the agent decides on itself.

• Python
• FastAPI
• ChromaDB
• bge-m3
• vLLM
• MCP
• Angular