Lab
Live demos, small tools and proofs of concept exploring AI, automation and productivity, and one that went all the way to production on Kubernetes. The experiments are rough by design — each one exists to answer a question. What holds up becomes a pattern on the AI page; what earns its place ships as a product.
Production demonstration
Java 25 · Spring Boot 4.1 · Angular 22 · Kubernetes
Unlike the experiments below, this one is finished and running: a full asset-maintenance application with Keycloak identity, PostgreSQL and object storage, deployed on K3s with Helm, Traefik and a Let's Encrypt certificate on a Hetzner server created from Terraform, with a public demo account. The engineering journey, in the order it happened:
New builds · autumn 2026
Each began as a question. Can one small server object per room referee a whole class playing at once? Can a single Python file wake a sleeping computer on time, every morning?
TypeScript · React 19 · Cloudflare Workers · Durable Objects
Live maths battles for classrooms and families. A teacher hosts, children join with a six-digit code, and everyone answers the same questions at the same time. One Durable Object per room holds the questions, the clock and the scores, so the server is the referee and nothing can be faked from a child's device. No database, no accounts.
Python · Tkinter · pygame-ce · macOS, Windows, Linux
A desktop alarm clock that plays an MP3 or your own recorded voice at a scheduled time, keeps the computer awake and registers a real wake from sleep with the operating system. Family schedules play a recorded message per event on a chosen speaker, and a PIN-protected phone remote on the home Wi-Fi stops, skips and records. One file, started with a double-click, and the whole folder is portable.
Experiments
Python · FAISS · OpenAI SDK
Local document search with chunking, MMR re-ranking and guardrails.
Agents · multi-tool orchestration
Scrapes roles, matches them against a profile and drafts outreach, with guardrails on every tool call.
Daily system · voice & text
A lightweight reflection routine for clarity, actions and tracking.
Learning in a handwritten style
A different way of taking notes with LLMs — five variations on the same idea.
From experiment to product
The Lab is the first step, not the last. The RAG and agent work here became the blueprint and production patterns on the AI page, and the same generate-and-review discipline built five products that are live today, took AssetCare to a live Kubernetes deployment, and most recently shipped BlitzMath.
Read how the experiments turn into practice, or go straight to what the practice produced.