Work
Open source,in operation,backed by tests.
Public repositories, self-built AI tooling and a full-stack case. None of it is a demo — it all runs, it's all readable and tested.
Open Source & AI Engineering.
2 public tools I built for my own work: lore (on PyPI, with CI and security scanning) and context-parachute (agent handoff across context limits). Both MIT-licensed.
My own Claude skills.
Self-built, run daily in my own operation.
Context management
Sessions survive context limits: handoff artefacts, persistent memory, orderly handovers between agents.
Cost-aware model routing
Each task goes to the cheapest model that still solves it cleanly — expensive models only where they earn their keep.
Quality gates
Nothing counts as done without evidence: verification before completion, review steps, guardrails against silent failures.
Engineering case
Pridovo
A World Cup prediction game built like production software. Next.js 16 and React 19 on the front end, Fastify 5 on the back end, PostgreSQL with Prisma. The API is covered by 273 tests, the critical flows additionally by Playwright end-to-end tests. Deployed via Docker and Traefik on my own server — live and operated, not a prototype.
It shows how I build software: type-safe, tested, cleanly deployable, transparently operated. How the same approach looks on the client side is in the example.
Infrastructure I run myself.
For development and agent workloads I run my own compute, reachable over a mesh VPN. Deployment goes through Cloudflare Pages and GitHub Actions, configuration as code rather than click-ops — the same discipline I use in client projects.
For a concrete conversation — [email protected]. Side projects and website experiments live on labs.semek.org.