MedPrüf

The problem

Foreign-trained doctors arriving in Austria face the Kenntnisprüfung — a German-language medical exam that decides whether their existing degree gets recognised. Three exam types: Kenntnisprüfung Wien (the broad clinical exam), KMP Innsbruck (a regional variant), and Pharmakologie (drug knowledge as a standalone). The existing prep market was either expensive private courses (€2K+) or scanned PDFs of old exams photocopied on a forum.

The architecture decisions that mattered

• Multi-exam from day one with one schema. Each question carries exam_type, topic, difficulty. A practice session is a query, not a separate codebase. Adding KMP Innsbruck after Wien was an evening, not a sprint.
• SM-2 spaced repetition on top of the standard quiz loop. The user sees what they don’t know, more often. Anki users feel at home; non-Anki users don’t notice the algorithm and just see “due today: 14 cards.”
• AI explanations are pre-generated, not live. 1,615 questions have written explanations stored in Postgres. The user never waits on a model call. The cost is a one-time bill at content authoring, not a per-user-per-session bill.
• Telegram Mini App + bot for re-engagement. The Mini App is the same Next.js code surfaced inside Telegram’s WebView. The bot pings users with “3 cards due today” at 19:00 Vienna time. Open rate: better than email by an order of magnitude in this audience.
• Daily quota counter for paid users. When AI explanations went paid-only in Apr 2026, I shipped a feature-flag-gated quota counter (AI_DAILY_QUOTA_PAID=15) so the kill switch is environment-only. No redeploy to disable AI in case of a model bill incident.

The pricing reset (Apr 2026)

Engineering is half the case study. The other half is whether the product earns — and what discipline went into the price grid. Live tiers as of Apr 2026:

• €27 first-3-months entry tier. Capped at one per account. Works out to ~€9/month — a thinner first ask for a market with crypto-payment friction and a long-tail conversion window.
• €19 single-month fallback. Demoted to fallback after the entry- tier reset; kept live for users who explicitly want monthly.
• €114 Founders’ / Cohort. Six months of full access — the lump sum that founders’ tier early-backers reach for.
• €29 / €79 / €149 / €229. Standard 1mo / 3mo / 6mo / 12mo tiers, waitlisted while Stripe Atlas (US LLC + EIN) clears. Crypto-only via NOWPayments in the meantime.
• Referral mechanic. €15 per starter / €25 per monthly+3mo / €50 per 6mo+12mo / €75 per Cohort. Stacks toward an exam-fee rebate when the referee passes — turns the existing cohort into a recruiting flywheel without paying paid-acquisition rates.

The Phase 1 reset shipped Apr 28, 2026 after a Researcher dispatch confirmed price wasn’t the conversion blocker — friction was. The discipline: test before rebuild. Phase 2 (€19/€49/€89/€149 across all tiers) is queued behind the Stripe Atlas approval and a survey of free users.

What I got wrong

The lesson

A medical exam-prep product is the inverse of a flashy AI demo: the model can’t hallucinate, the spaced-repetition math has to be reproducible, and the German translations have to be exact. The AI features are scaffolding around a question bank that has to be right. If the question is wrong, no amount of GPT-4 can save you. Most of the engineering went into the boring layer: editorial tools, the duplicate-question detector, the per-exam-type scoring rubric.