Doctolingo

Pure SwiftUI app, iOS 17+. No backend in V1 — everything runs on-device or against Azure's REST endpoint. Audio capture uses AVAudioEngine with a single .playAndRecord session activated lazily at point-of-use (activating in App.init() fails on real devices — UIApplication lifecycle isn't ready). Scored attempts persist to a local SQLite DB via GRDB with UUID primary keys, created_at / updated_at / deleted_at columns, and user_id = 'local' until Clerk auth ships in V2.

Two scorers run per recording. On-device (Apple): SFSpeechRecognizer transcribes the attempt; a custom Castilian G2P (SpanishG2P.swift, ~25 rules — ch→tʃ, ll→ʝ, rr→r, qu/gue/gui, güe/güi→ɡw, ce/ci→θ, ge/gi→x, ñ→ɲ, word-initial r vs flap ɾ) converts both reference and transcript to IPA, then computes (1 − PER) × 100 via phoneme Levenshtein. Free, instant, offline. Cloud (Azure): The WAV is sent to Azure's REST Pronunciation Assessment endpoint; response is decoded with flat-key custom CodingKeys (Azure's REST shape differs from the SDK's nested shape — AccuracyScore, FluencyScore, PronScore sit directly on each NBest entry, not under a PronunciationAssessment sub-object). Both scores are stored per attempt to enable a later correlation analysis — the goal is to phase out paid Azure once SFSpeech+G2P correlates tightly enough.

Reference audio is pre-generated offline via voicebox.sh (edge-tts + Microsoft neural Spanish voices like es-ES-AlvaroNeural), bundled in the app's content/es/audio/ directory as ~200 WAVs. No per-user TTS cost; no network required after install. Each phrase carries an id, en, es, and category. V1 ships 3 scenarios × ~50 phrases each: chest-pain intake, abdominal-pain intake, headache intake — all written against USMLE OPQRST vignettes (not Canopy phrases, to avoid any derivative-content argument).

Simplified Leitner with 3 boxes — Box 1 (due +1 day), Box 2 (+3 days), Box 3 mastered (+30 days). Scoring ≥ 80 promotes to the next box; < 80 drops back to Box 1. Persisted as JSON in UserDefaults (demo-scale). StreakStore tracks current streak + today's phrase count + daily goal; streak increments when last-practice was yesterday, resets after a full day missed, normalizes on foreground. Haptic feedback (UIImpactFeedbackGenerator) fires on score arrival with style .medium on pass, .rigid on fail — tactile delta before the user reads the number.

Build number is derived from git rev-list --count HEAD via scripts/bump.sh so every commit that ships is monotonically higher in TestFlight. PerfLog.swift times every hot path (audio session activation, AVAudioPlayer init, reference-audio bundle lookups, Azure roundtrip) — critical for catching Debug-build slowness, since Swift's synthesized Codable decodes JSON ~50–100× slower in Debug than Release and can starve the main thread at launch. Reference-audio URL lookups are cached per phrase (150 phrases × many SwiftUI re-renders × recursive bundle walks added up to visible lag).

SFSpeech transcripts never leave the device. Recorded WAVs are sent to Azure solely for scoring and are not persisted by Microsoft per their Speech Service retention policy. The app stores attempts (transcripts, scores, timestamps) locally in SQLite at Application Support/Doctolingo/doctolingo.sqlite — explicitly outside Documents/ so recordings don't surface in iTunes file sharing. No analytics, no crash reporting, no third-party SDKs in V1.

No Android. No free-translation drills, grammar exercises, or vocabulary flashcards — V1 is pronunciation-only. No social features, leaderboards, or friends. No AI-patient role-play (captured as V2). No Mandarin, Arabic, Vietnamese — Spanish-only ship. No Supabase sync, no RevenueCat subscription gating, no voice profile enrollment. Scope discipline is the feature; Canopy is enterprise-bloated and Duolingo teaches you to order coffee, not to ask about chest pain character.