I have an ordinary full-time job and a train ride to it. This page is about the other half — the products I build on evenings and weekends, with an AI agent working next to me and a single small server doing the hosting.
None of it is a mockup. Every link below points at something running right now, with real users on the other end.
Live right now
-
Lingou Livelingou.de
Real-time translation for churches, conferences and events. The speaker talks; everyone in the room follows along on their own phone, in the language they picked — eleven of them, from German to Ukrainian to Arabic. No app to install, no headsets to hand out, no hardware. A QR code on the wall.
-
Lingoulingou.de/aprender
The app the translation grew out of: German for Brazilians. Flashcards, grammar, themed modules — and a voice tutor you can hold an actual conversation with, which is the part that finally made the vocabulary stick for me.
-
black-hole.studioavatar.black-hole.studio
My studio domain, and where the video work lives. Avatar takes one photo and fifteen seconds of your voice, then turns anything you type into a video of you saying it. Stories is the pipeline behind it: script, voice, footage, render, publish — priced per second, because I got tired of being surprised by invoices.
-
Helpyhelpy.bot
A calm Mac assistant for people who find computers stressful. It listens, looks at what's on screen, and points at the one thing to click next — never doing it behind your back. I built it thinking about my parents.
-
ClaudeClawopen source
A coding agent that lives in my Telegram. I describe a change from the train, it does the work on the server, runs the tests and sends back the diff or the finished video. Most of the projects on this page were partly built this way.
How the night shift actually runs
One server, many products
Every service above shares a single modest VPS behind Cloudflare tunnels, with backups before each deploy and a watchdog that wakes an agent to fix things when a site goes down at 3am. Constraints make better engineers than budgets do.
AI as a colleague, not autocomplete
I don't ask a model for snippets. I give it the repo, the tests, a scope document it has to keep honest, and the ability to run what it wrote. My job shifted from typing to deciding, reviewing, and saying no.
Shipped beats finished
Lingou Live went in front of a real congregation before it was pretty. Avatar was billed against a real invoice before it had a landing page. Nothing here was designed in a doc and then abandoned.
Before this
Years of ordinary, useful product work: React and Next.js front-ends, Flutter apps, .NET and GeneXus back-ends, a VS Code extension a few thousand people installed, a Medium clone with Stripe subscriptions built to learn Jamstack properly. Those older projects still live on GitHub.
TypeScript · React · Next.js · Node · Flutter · .NET · PostgreSQL · SQLite · Docker · Cloudflare · Playwright · FFmpeg