Why an education app is about completion, not content

Most education apps ship beautiful course catalogs and terrible completion rates. A user downloads, watches lesson one, and never returns. AI helps in one specific place: generating a short quiz at the end of each lesson so the user has a reason to mark the lesson complete. Without that loop, completion rates sit under 10% in most categories.

I have not shipped an LMS, but I have shipped a video-and-quiz loop inside another app, and quiz generation from lesson transcripts is cheap and reliable when the transcript is chunked right.

What you actually need to build

• Course list + detail: Courses have lessons. Lessons have a video or text body and, optionally, a quiz. Start with 3-5 seed courses.
• Video playback: Embed HLS video. Track resume position per user per lesson. 0.8x to 2.0x speed control.
• AI quiz per lesson: 3 multiple-choice questions generated once per lesson transcript, cached globally. Pass threshold is 2 of 3 correct to mark as complete.
• Progress tracking: Percent complete per course, streak of lessons completed, next-lesson suggestion.
• Certificate on course complete: A simple PDF with the course name and user name. Cheap retention win and a social share hook.

No live classes, no instructor dashboard, no payments in v1. Course payment is its own build and comes after you know which courses retain.

The stack I use

• React Native + Expo.
• expo-video for HLS playback with resume position.
• Supabase — courses, lessons, user progress, quiz cache.
• NestJS — quiz generation endpoint. Given a lesson transcript, generate 3 MCQs in structured output. Cached on lesson_id so the same lesson never generates twice.
• react-native-pdf-lib for certificates.
• Claude Code + 11 AI agents — scaffold the course list, lesson player, and quiz screens.

Real build time

With the boilerplate, 4 weekends.

• Course schema + import pipeline: ~6 hours.
• Video player + resume tracking + speed: ~8 hours.
• Quiz generation endpoint + cache: ~5 hours.
• Quiz UI + scoring + retry flow: ~5 hours.
• Progress + certificate PDF: ~5 hours.
• Store submission: ~4 hours.

About 33 hours.

Where people get stuck

• Video buffering on mobile networks: Ship HLS, not MP4. A mobile user on 4G will not wait for a 50MB MP4 to preload. HLS adapts and buffering drops to near zero on the 5 test devices I ran this on.
• Quiz quality from long transcripts: Feeding a 20-minute transcript to an LLM gives vague questions. Chunk to 3-minute segments, generate one question per chunk, then pick the best 3 by a simple relevance score. Quality jumps hard.
• Completion gaming: Without a time-on-video check, users skip to the end and get a certificate. Enforce server-side that video_watched_seconds >= lesson_length * 0.8 before the quiz unlocks. Do not trust the client.

Skip the setup

HLS playback scaffold, Supabase schema with progress tracking, cached quiz generation endpoint, certificate PDF template — pre-wired. The 11 AI agents scaffold the player and quiz screens from a prompt.

See pricing