Goal — A translator between German, English and Bavarian, in both directions. The benefit is that a dialect nobody builds a tool for becomes translatable at all — instead of being ironed out as faulty German.

Challenge — A dialect without a binding orthography is the most awkward case a translation tool can face: there is no single correct target form to check against. The same tools that work reliably between German and English quietly correct Bavarian away.

Current state — The public page is live and reachable over an encrypted connection. The key handling is what carries it: the key to the language model sits on the server and never in the browser — the page cannot read it out because it never receives it. That is the construction, not a precaution added afterwards. The front is public; the full interface sits behind a password. And the public page names no model, no institution and no partner. in progress

What part AI plays — In the product: generation, and it is the actual core. A language model produces the translation; the surrounding application decides nothing about the text, it holds the key and passes the request on. In the building: the application itself, the key handling and the construction of the page are developed by AI agents, with human approval.