An AI German-conversation tutor for English speakers, from absolute beginner to B1. A pnpm and Turborepo workspace: two Next.js apps and a Node worker over shared packages, TypeScript throughout, Supabase and OpenRouter.
You talk to it in German. It replies at your level, corrects you in the way your level can use, and keeps track of what you know — seventy mini-levels from a first hallo to holding an opinion. When four in five of your recent messages come back error-free, it moves you up on its own; say weiter (or press the button) to move on early.
What you say is never stored. A conversation lives in your browser and dies with the tab; the database keeps only what the practice meant — your mistakes, your vocabulary, and a per-level score.
A tutor that talks (/chat). German conversation at your level, streamed, with corrections your level can use, retrieval over a hand-authored knowledge corpus, and long-term memory of your goals and interests. Speak to it and have it read back to you; the speech model first, your device's own voice as a real fallback.
A curriculum, not a model's opinion (/curriculum). Seventy mini-levels, 420 words and 86 grammar patterns as data. Progression, mastery, review scheduling and recommendations are pure functions over your own record.
A trainer that runs the drills (/trainer). A tool-using agent that opens a session with an agenda — due words, recurring mistakes, the level's grammar and culture notes, today's simple-German news — quizzes one item at a time, judges the answer, records the outcome through the deterministic review engine, and shows every tool it used.
A trainer that learns from you. Rate any reply, or just ask for a change, and closed-set dials move: how much it explains, how hard it pushes, how long a sitting runs, what the agenda leans towards, whether it ends on the news. Disputed corrections make it check a rule before it corrects. Everything it learned is on the settings page with a button that forgets it — next to a choice that button leaves alone: a friendly trainer or a strict one.
Scenarios and comprehensions (/scenarios). Fifty practice items found in your own words by vector search: dialogues played as roleplay, passages answered as a quiz.
Review and progress (/review, /dashboard). What is due, what is weak, what keeps going wrong, and the per-level score that decides when you move.
An operator app of its own (apps/ops). On its own origin, behind a password and an authenticator code: euro spend by task and model, and the AI Product Ops board described below — which opens every scan with what learners said about the trainer. It records jobs; a worker runs them.
