The need
Practise our class material between lessons.
Kory Kekewich← All projects
Kanrabat
From classroom notes to daily French practice.
A Canadian French practice app built for my class and used by my classmates.
At a glance
What I built
Exercises, scheduled review, and a contextual AI tutor.
My role
Product design, React development, grading, and AI integration.
Built with
- React + Vite
- Phone-first learning interface.
- Node.js + Vercel
- Serverless tutor routes and session checks.
- Anthropic API
- Lesson-aware explanations.
- localStorage
- On-device learning history and progress.
- Web Speech API
- French audio playback, voice availability permitting.
- Node checks
- Grading, identity, and review tests.
Product preview
App screenshots from Niveau 2 (Level 2), shown in phone frames.
Engineering decisions
Fair answer checkingI accept authored alternatives and check verb forms before forgiving a typo.
- Challenge
- Exact matching rejects valid French. Loose typo matching can accept the wrong verb form.
- Outcome
- The grading checks reject 84 tested verb-form confusions while preserving the class's accepted alternatives.
Review that remembersI gave learning items distinct identities and used their attempt history to build a staged review queue.
- Challenge
- The same sentence can appear in different exercises or carry different meanings. Its learning history must not get mixed up.
- Outcome
- Practice feeds an on-device review schedule. Identity and scheduling checks cover the logic; review intervals remain provisional.
A tutor with lesson contextI send the exercise, both answers, and the relevant rule to Claude through a server-side API with session checks and request limits.
- Challenge
- A general chatbot does not know which answer a learner gave or what their class has taught.
- Outcome
- Learners can ask why an answer was wrong without rebuilding the context themselves. AI explanations can still make mistakes.
How AI is used
Explanations with context
The tutor uses the Anthropic Messages API with authored lesson facts and conversation history. It is contextual prompting, not retrieval-augmented generation or an autonomous agent. Server-side sessions, request limits, and token limits constrain access and usage, but do not guarantee a spending ceiling or error-free answers.
My contribution
I designed and developed Kanrabat around my own Canadian French course, covering the interface, curriculum data, grading, review scheduling, server-side AI integration, regression checks, and deployment. Development included AI-assisted coding tools.
Where it stands
Used by my classmates and me.
Continuing to improve alongside our Canadian French course.
Next step
Keep developing the app alongside my own French learning, with classroom use informing what comes next.