AI Study Coach

The architectural problem is AI cost: a free-tier study app that pays for every generation against a per-request OpenAI bill goes bankrupt by the second power-user. The backend treats AI calls as billable resources to defend. Every generation hashes its input (SHA-256 of user, type and content) and serves identical requests from cache without re-billing OpenAI. Every client request carries an idempotency UUID, so a network retry never double-counts against the user's quota. The AI is instructed to return JSON only, validated against Pydantic schemas, with one repair-retry on malformed output before failing gracefully. On the device, Room is the source of truth: AI-generated study sets persist locally the moment they come back, so the app is fully usable offline once a generation completes. The auth flow is guest-first — users can start immediately with no sign-up and upgrade later, with all data preserved via account linking. Custom OTP verification (branded HTML through Resend, 10-minute expiry, 60-second resend cooldown) replaces Firebase's generic email so users never leave the app's visual world.

A study app with AI inside has two failure modes that bankrupt it. The first is treating AI as a free resource and watching one power user burn through a month of margin in an afternoon. The second is throttling so hard that nobody sees the value before they bounce. AI Study Coach threads that needle with a native Android client and a FastAPI backend that treats every AI call as a billable resource worth defending.

Why native Android, not React Native

The product surface is dense: a dashboard with stats, study-set list and detail screens, a flashcard reviewer with flip animations and SM-2 scheduling, a quiz engine with four question formats and a chat tutor, plus folders, analytics, a Pomodoro timer, notifications, social leaderboards and an admin portal. Going native in Kotlin and Jetpack Compose with Material 3 gave dynamic colours on Android 12+, a single typed codebase end to end (Kotlin on the device, TypeScript-shaped Pydantic schemas on the backend), and direct access to Android subsystems: AlarmManager + BootReceiver for streak reminders that survive reboots, ML Kit Text Recognition for OCR import, EncryptedSharedPreferences for token storage, and a foreground service for the Pomodoro timer that keeps running while the app is backgrounded.

AI economics, designed in

The FastAPI backend treats AI as a metered resource. Every generation request hashes its input — SHA-256 of (user_id, content_type, content) — and serves a cache hit when the same content rolls back through. Identical re-runs cost nothing and don't count against the user's quota. Every client request carries an idempotency UUID, so a network retry never double-charges. The AI is told to return JSON only, validated against Pydantic schemas, with one repair-retry on malformed output before a graceful failure. Quota tiers are enforced at the API layer (1/day anonymous, 10/day free, unlimited premium) alongside a five-per-minute rate limit. The result: AI spend is bounded per-user and per-day regardless of how aggressive the usage pattern gets.

Spaced repetition done properly

Flashcards run on the SM-2 algorithm, the original spaced-repetition schedule. Each card carries a difficulty, an ease factor and an interval in days. After a review, the algorithm computes the next interval based on how the user rated their recall; cards converge to a stable monthly cadence once known and resurface as "due" automatically. The reviewer surfaces this honestly: it shows the current interval (days / weeks / months), the known-vs-review breakdown on completion, and lets the user fall back to an "All cards" mode for cramming when SRS feels too slow.

Guest-first, then upgrade without losing anything

Sign-up friction kills onboarding for a consumer app. Users start immediately as anonymous guests with a single AI generation a day, full access to flashcards, quizzes and study plans, and zero email field. When they're ready to upgrade — usually because they hit the quota or want their study sets backed up to the cloud — account linking moves their existing data onto the registered account without re-creating anything. Email sign-ups go through a custom 6-digit OTP verification flow with branded HTML through Resend instead of Firebase's generic verification email, so the visual world never breaks. Google sign-ins are auto-verified.

Offline by default, sync when online

Room is the source of truth on the device. AI-generated content lands in local storage immediately, so the app is fully usable offline once a generation has completed. For registered users, study sets sync to the backend in the background, and a fresh install signs in and restores everything from the cloud. The chat tutor is a deliberate exception — it needs network for streaming responses — but the rest of the app keeps working at 35,000 feet.

What shipped

A native Android app on Play with a viable free-tier business model underneath: guest-first onboarding, SM-2 flashcards, four quiz formats, dynamic study plans, an AI chat tutor scoped per study set, OCR import from camera or gallery, PDF and URL ingestion, gamification, analytics, a Pomodoro timer that runs as a foreground service, social leaderboards and a server-rendered admin portal. The Next.js marketing site at aicoachstudy.com and the FastAPI backend deploy together from one Docker image on Render.