← All projects

AI training · English tutoring

Build an AI tutor for English practice

Try lesson planning, turn-by-turn roleplay, targeted feedback and vocabulary practice.

48bilingual training examples
23/24correct held-out routes
0.977held-out macro F1

Try the English tutor

Ask in English or Portuguese. Use the controls to set the topic, difficulty and lesson length.

Choose an example or enter a request to begin.

Four teaching workflows

  • Lesson planning: timed objectives, a model, guided practice, independent production and a success check.
  • Conversation practice: scenario-based turns, repetition, sentence starters and bundled audio playback.
  • Corrective feedback: targeted examples for third-person agreement and past tense, with a retry.
  • Retrieval practice: an answerable drill, delayed feedback and a follow-up task.

Training and evaluation

Train a multinomial Naive Bayes classifier on labeled EN/PT-BR requests and export its weights for browser inference. A small score margin routes uncertain requests to a clarification. The tutor prompt pack and eight instruction-response examples specify teaching behavior and acceptance cases for a future generative system.

Held-out examples by expected and predicted intent
Expected intentLessonRoleplayFeedbackPracticeClarify
Lesson60000
Roleplay06000
Feedback00600
Practice00051

The one missed request was a vocabulary drill that triggered clarification. The small, authored holdout measures intent routing only. Broader evaluation would use separately authored requests, multi-intent inputs and review of lesson quality by teachers.