Challenge
LangBoost's existing language-learning app corrected grammar but didn't teach. Students drilled flashcards, hit a plateau around week three, and churned. The founder had 4 weeks and a fixed budget before an investor demo - no runway for a multi-month rewrite
Discovery
We ran a 3-day workshop with the founder and one target user. The real job was not "corrections" but conversation practice with a patient tutor - the kind of practice a $60/hour human tutor provides. Flashcards were a side effect, not the product
That reframing unlocked the architecture: tutor first, drills second. Spaced-repetition scheduling now feeds the tutor conversations instead of being the whole app
What we shipped
AI tutor
- Claude Sonnet system prompts tuned per CEFR level (A1-C1) with grammar-aware feedback
- Live conversation transcripts saved, re-indexed, and folded back into each learner's next session
- Pronunciation scoring via GPT-4o audio input - no third-party SDK required
Adaptive drills
- SM-2 spaced-repetition engine that pulls cards from the week's conversation errors
- Difficulty tuning per learner - cards the tutor catches repeatedly surface more often
Commercial surface
- Stripe subscriptions with 7-day free trial and annual/monthly plans
- Admin analytics: DAU, tutor-session length, plan-level retention
- Student dashboard showing streaks, level progress, and upcoming cards
Why this stack
Rails + Hotwire let the two engineers ship a multi-screen dashboard without a separate front-end build. PostgreSQL JSONB stored tutor state - spec + transcript + error log - without premature schema. Claude Sonnet over GPT-4 for the tutor persona (more patient, less hallucinatory in correction scenarios); GPT-4o kept for audio transcription. Heroku + Stripe meant one deployment to worry about on demo day
Outcome
- Shipped in 4 weeks - investor demo made, term sheet signed two weeks later
- 5.0 Upwork review on first engagement
- +22% month-2 retention once the tutor-led flow replaced the old flashcard funnel
"Aimeice delivered in four weeks what our last vendor couldn't ship in four months. Straight talk, real code, zero hand-holding required"