Work
SaaS · AI

NorskCoach.

A Norwegian-language coach that finds your weak spots in two minutes, then talks with you. I treat teaching quality as an engineering problem: a tutor has to be correct before it gets to be convincing.

norskcoach.no
Next.jsSupabaseGeminiClaudeFSRS
NorskCoach. — the live site
Role
Sole engineer, design → ops
Type
Learning product with AI
Status
In development, live users
Focus
Model-quality governance

The situation

Most language apps drill you on what they already have. Working life in Norway needs more than the textbook: ordering a coffee, following colleagues, holding an everyday conversation. NorskCoach starts from the learner instead of the syllabus: a diagnostic that takes two minutes, needs no account, and finds the gaps before any lesson is chosen.

What I made

A diagnostic onboarding that maps a learner's weaknesses, a practice loop that schedules what to revisit with spaced repetition, and a voice partner you tap and talk Norwegian with. Behind it sits an admin surface for content and oversight, so what the coach says can be reviewed rather than trusted.

The product is live with real learners while I build it. The teaching-quality work is the hardest part and it is not finished.

Three decisions

  1. Two models, tiered by what can go wrong. Conversation practice and vocabulary drills run on a fast model. Grammar explanation and correction of a learner's errors escalate to a stronger one. Escalation is a design decision, not a fallback: correctness-sensitive tasks never run on the cheapest model.

  2. Spaced repetition owns the schedule. What comes back, and when, is decided by an FSRS scheduler over the learner's own record, not by the model's mood. The AI teaches; the algorithm remembers.

  3. Correct before convincing. A tutor that teaches confidently wrong grammar is worse than no tutor. Model selection is treated as a curriculum decision, and the admin surface exists so a human can see what was taught.

Outcome

Live at norskcoach.no with a free, no-account diagnostic and a voice partner. Still open: the quality bar for spoken correction, and how far a fast model can be trusted before the stronger one has to step in.

If you are building something similar, I read every brief myself and reply within 48 hours.

Project brief