AI-Powered English Learning App
Aligned with CEFR levels A1-B2, with three modules: handwriting OCR scoring, AI speech scoring, and AI conversation, backed by a 7,800-word vocabulary database.
Duration
In development (closed beta)
Team
2-3 people
01 // OVERVIEW
Project Overview
An English learning app for Japanese learners, supporting CEFR levels A1-B2 with three modules: writing, speaking, and AI conversation. Automatic scoring and feedback are implemented via the Google Gemini API. It is currently being used by 30 students in a closed beta at a specific tutoring school.
02 // KEY FEATURES
Key Features
Writing Assessment
Students photograph a handwritten English essay with their smartphone; the text is extracted via OCR and automatically scored by the Gemini API. Content, structure, vocabulary, and grammar are each scored out of 4 points, with specific explanations of corrections and vocabulary analysis also provided.
Speaking Assessment
Records about 60 seconds of speech, transcribes it with Whisper, and scores it with the Gemini API across four criteria - content, vocabulary, grammar, and naturalness - generating improvement suggestions.
AI Conversation
Learners select a CEFR level and number of turns, then have a real-time voice conversation with the AI. The chat-style UI allows the AI's responses to be played back as audio.
CEFR-J Vocabulary Database
Manages a list of 7,800 words in MySQL, filtered by level and injected into the scoring prompt to ensure evaluation accuracy.
03 // TECHNICAL DEEP-DIVE
Technical Highlights
Solution
Injecting the Vocabulary Database into Prompts
Many AI learning apps leave scoring entirely to the AI, but this app retrieves level-appropriate vocabulary from the database and explicitly includes it in the prompt, achieving systematic and consistent scoring.
Solution
Multi-Dimensional Feedback
Rather than simple scoring, the app automatically generates feedback learners can actually act on, including specific explanations for each correction, vocabulary-level classification, and identification of common error patterns.
04 // IMPACT & RESULTS
Results
30
Test Users
7,800
Words in Database
3
Modules: Writing, Speaking, AI Conversation