- Problem
- German place names have a recognisable structure, but no useful set of rules for inventing them.
- Constraint
- The model had to run entirely in the browser: no backend, no account, no inference bill.
- Decision
- I trained a region-conditioned character GRU on 10,729 cleaned Destatis names and ported its inference from PyTorch to dependency-free JavaScript.
- Result
- A 1.3 MB game with cross-runtime tests and 31% regional accuracy against a 7.7% random baseline. The limitations are published too.
Most of what I know started with getting it wrong in public.