README.md
Anonymized case study: client, data and code are under NDA.
The problem
Preparing for a certified language exam takes many realistic papers, and writing them by hand is slow and expensive.
What it does
- Four modules: reading, writing, listening and speaking, the last one with a voice engine.
- Generation from the dashboard: pick the model, every cost is recorded and the price per paper is always known.
- Grading with feedback in the style of a tutor, reviewed by a language expert.
Constraints
Same structure as the real exam, but original interface, images and content: official material never enters the project.
Now
Reading, writing and listening are heading towards stability, speaking is under construction. The engine that generates papers runs as workers and shares a common dictionary, hints and per-document grading through a cache, so each paper costs less.