AI Flashcard System

Card Generation and Spaced Repetition

Design a flashcard app that turns a student's notes and PDFs into cards with a language model, then schedules each card for review just before it would be forgotten, on and offline.

Overview

A flashcard app shows a question, the student tries to recall the answer, then rates how well they did. Spaced repetition uses those ratings to schedule each card: a card you know well comes back in weeks, one you missed comes back tomorrow. Anki and SuperMemo are built on this; SuperMemo's SM-2 algorithm dates from its first program in December 1987.

The AI part removes the slowest step for students: writing cards. The student uploads lecture notes or a textbook chapter, and a language model drafts question and answer pairs. Drafts need care. A model can write a card whose answer is not in the source, or ten cards that ask the same thing, so every card is checked against its source text and the student approves the deck before studying it.

Questions to ask first

How many learners? — 1 million a day, about 60 reviews each
What goes in? — PDFs, pasted notes and web pages; 50,000 documents a day, about 30 pages each
Must review work offline? — Yes, on phones; sync when back online
Which scheduler? — FSRS by default, SM-2 available; both run on the device
Do AI cards go straight into the deck? — No, the student approves drafts first
Interoperability? — Export decks as Anki packages

What you will be able to do

Build a pipeline that drafts cards from a long document and rejects ones its source does not support.
Explain SM-2 step by step and trace a card through several reviews.
Explain FSRS's forgetting curve and how desired retention sets the next interval.
Design offline review so two devices never corrupt a card's schedule.
Size the system and say which parts need a server at all.

Key concepts

Spaced repetition
Forgetting curve
SM-2 ease factor
FSRS and desired retention
Chunked generation
Grounding check
Review log as events
Offline-first sync
Idempotent upload

Functional requirements

Upload a PDF, notes or a web page and receive draft cards (question and answer, or cloze deletion)
Review, edit, approve or discard drafts before they join a deck
Show the cards due today and schedule each card from the student's rating
Review offline and sync across phone, tablet and web
Show progress: due counts, retention, a forecast of upcoming reviews
Export a deck as an Anki package (.apkg)

Non-functional requirements

Drafts for a 30-page document ready in under 3 minutes at p95 (a target we chose)
The next card appears in under 100 ms, because it comes from the local database (a target we chose)
No review is ever lost or counted twice, whatever the order devices sync in
The same review history gives the same schedule on every device
Generated cards cite the passage they came from

Interview questions

Walk me through SM-2.
What does FSRS do better than SM-2?
How do you support offline review on several devices without conflicts?
How do you keep AI-generated cards correct?
How would you process a 500-page textbook?
Where does the server matter, if scheduling is on the device?