Prototype
KI-Sprachtrainer
A personal language coach where you simply talk freely – the learning logic runs invisibly in the background.
99.9%
cache hit rate in conversation
01 · Starting point
Many learners understand a language passively but stall when speaking. Classic apps train vocabulary and gap-fill exercises – not free conversation.
02 · Solution
- 1 Speech-to-speech using a real-time model, chosen after a direct comparison of two providers.
- 2 Three tempos: code runs continuously, a small model regularly feeds stage directions into the live session, a strong model analyses afterwards.
- 3 Usable as a progressive web app directly in the browser and on a smartphone.
03 · What I learned
Models can be too helpful
Newer transcription models silently “correct” learners’ mistakes – exactly the mistakes the coach needs to see. Model choice depends on the use case, not on the leaderboard.
Cost is architecture
Whether an AI application is economical is decided by session length, caching and model choice – not by the invoice at the end of the month.