AI engineer · Prague · at Medevio since 2026
At Medevio, a platform Czech and Slovak clinics use to talk to their patients, I build AI features that save doctors and nurses time on admin. I got there the long way round: QA, built and led a QA team, then release engineering, then full-stack development and AI product features, which is why I’d rather measure a feature than admire it.
These days most of the code I ship is written by coding agents. I set the rules they work under, read what comes out, and measure whether it helped.
Most of the projects below run on my Raspberry Pi, and so do some of my agents. I can work from wherever my phone is: I send an agent a task, it works on the Pi, and I check the result later.
Medevio
I own the AI parts of the product: drafted replies for doctors, summaries of patient requests in Czech and Slovak, processing of calls and text messages, transcripts of meetings. Prompt, code, rollout, and the evaluation that decides whether it ships at all.
I keep the conventions, skills and runbooks the agents work from, with credentials cut so an agent can run a data job but can’t touch a live clinic. Most of my attention goes to evaluation: judges with repeated votes, holdout sets, significance tests. Changes that don’t survive the numbers get dropped, including a few of mine.
Real, running, and mine. Much of it lives on a Raspberry Pi in my flat. The line under each one says what the build script found when it checked the link.
Huginn
One inbox for Slack, Gmail, ClickUp, GitLab and Signal. I was tired of five apps asking for attention, so I wrote one dashboard where I can reply, draft, react or mark things done, and it learns from what I mark as spam or important. It started on the Pi and is now a Mac app, with tokens encrypted at rest.
Jev model router for Claude Code
A small proxy that picks the model for each Claude Code turn: cheap ones for cheap turns, the expensive one when it matters, every decision logged. It exists because I looked at my bill.
Realiterier
A watchdog for Czech property listings. It pulls ads from four portals, reads the text with an LLM for things the portals don’t filter on (pets allowed, no commission, brick or panel), groups the same flat across portals, flags the usual scam patterns and mails you within minutes. Right now only the web part runs; the scraper stays off until I decide what to do with it.
kviff.ai
A chat guide to the Karlovy Vary film festival. It crawls the programme, pulls film data from TMDB, OMDB and ČSFD, and answers questions like what fits between two screenings. It wakes up for the festival in July and sleeps the rest of the year, which is also when I stop paying for it.
Deskohraní
Sign-ups for the board-game club I run on Thursdays: a form behind a home-made dice captcha, a capacity that closes the form on its own, attendance, articles, CSV export, nightly backups. Node 22 with nothing but the standard library, in Docker behind a Cloudflare tunnel.
Žižkostel
A new website for a community centre in Prague 3, built so its editors can publish events without calling a developer: Payload CMS with a Czech admin, drafts and versions, two languages with Czech as the fallback. The client is still writing the content; the demo runs on my domain.
The Pi itself
A Raspberry Pi 5 in my flat hosts the sites above, plus a Signal assistant with local Czech speech-to-text, a private family app, and nightly encrypted backups to S3. A watchdog speaks up only when a backup goes stale, which is the one kind of monitoring I trust.
ABUGO Group
AI automation engineer. Internal AI agents in n8n, Make and TypeScript.
- An AI-policy chatbot in Slack that employees used every day to check whether a given AI tool was allowed under the company’s policy.
- A domain-scan agent that checks foreign websites as possible customers for group companies like Smartsupp, and flags dangerous, fraudulent sites based on user input. Hundreds of checks a day.
Make (Celonis)
Developer experience engineer. Automated the release pipelines, set up bug-ownership agreements between teams, helped build automated instance deployments, and drove the company-wide push to test earlier.
- Put AI into the triage of support tickets on their way to development: hundreds a week.
Make (Celonis)
QA lead. Quality for a complex no-code automation product, with automated tests wired into the deployment process.
- Built and led a QA team of six.
- End-to-end tests in Cypress that caught dozens of critical bugs before they reached production.
- Cut the automated test run from 7 hours to 20 minutes by running it in parallel.
Rohlík
QA engineer. Manual testing and test analysis for an online grocer, which I started automating at the API level.