Progress needs evidence
Level estimates use independent, assessable work. Assisted answers and deliberate reviews are excluded from promotion evidence, so using help does not falsely inflate progress.
01 / Applied AI · LearningLive product
French Practice Studio
A language-learning system built around the learner.
An adaptive French practice workspace covering reading, listening, speaking and writing. Built from a problem I know personally: making consistent, useful progress in another language.
Open live productFrench Practice StudioFR ↗
Un peu, chaque jour.
A little practice.
A world of possibility.
A world of possibility.
Read
Listen
Speak
Write
Practice → Feedback → Progressà vous.
Product illustration · Explore the actual interface through the live product link.
Start with the problem.
A learner needs more than an endless feed of exercises. Practice needs to reflect their level, preserve their work, and make progress visible without mistaking assisted answers for independent ability.
My contribution
Personal product: built and iterated the application, adaptive practice rules, AI integrations, account experience and deployment.
- Keep each learner’s progress private.
- Make generated exercises assessable and avoid repeated content.
- Handle speech and AI failures without inventing results.
- Control burst traffic and paid provider usage.
Follow the flow.
Learner
Next.js / Vercel
Supabase Auth + RLS
AI exercise & assessment
Progress evidence
A deliberately simplified view of the product. The decisions below explain the boundaries that matter.
The decisions behind the interface.
Privacy at the database boundary
Verified sessions and per-user Row Level Security protect progress. Server routes mediate paid AI calls; browser clients never receive privileged credentials.
Bounded AI, explicit failure
Generated exercises are validated with bounded retries. Rate limits are backed by shared database state. Provider failures appear as failures rather than fabricated scores.
What exists. What remains open.
- A deployed product with public registration and a private learner workspace.
- Tracked checks cover assessment rules, novelty heuristics and generated-task validation.
- Branded verification and recovery emails were verified through Resend in the project release record.
Known limitations
- CEFR estimates are study heuristics, not an official language certification.
- Semantic novelty is heuristic; different wording does not guarantee a different learning objective.
- Microphone behavior depends on the device and browser; complete device coverage is not claimed.
- No public adoption or learning-outcome study has been published.
Keep asking better questions.
Publish a versioned evaluation set for exercise quality and a repeatable speaking-assessment comparison.
Next project / Live