The flashcard app that shows its work.
A calibrated, source-validated study app for data science, ML, and AI. Every card carries a difficulty rating that learns from real answers. Every claim is checked against primary sources. Pick a Field, open every domain, and climb a ranked Arena against people at your level.
NAN · PEEVAL · GENERALIZERO
Cards rated like chess players, a mastery map that moves with every answer, and a ranked Arena. Real screens from the app.
Most flashcard apps publish a deck and call it done. We're building infrastructure — and the only way to earn trust before launch is to describe it before measuring it. Numbers come later, on the quality report.
Every card carries an EigenScore — an ELO-style difficulty rating on the same 800–2400 scale chess uses. Your own EigenScore moves with every answer. Combined, the two predict your odds of getting a card right before you answer. Cards whose observed difficulty drifts from prediction get reviewed.
Every factual claim in every card is extracted, web-searched against a per-track allowlist of authoritative sources, and entailment-checked against the retrieved passages. If a claim can't be supported, the card doesn't ship.
Each domain ships with a published maturity tier — Seeded, Drafting, Live, Calibrated, Mature — and we won't pretend we're further along than we are. Users can flag any card in one tap; a flag queues the card for that night's review pass (it stays visible) — the review clears it, repairs it, or pulls it, in public. Flagging never hides a card on its own.
Six gates between a draft and your queue. A card that fails any gate is rewritten or dropped — drafting state never reaches you.
Calibrate once, then each sprint is a deliberate blend — what's due, what's shaky, what's next, and what you haven't seen in a while. Every card is pulled close to your level, so you're stretched, not steamrolled.
Spaced repetition, FSRS-style. A card comes back exactly when the model thinks your recall is about to dip below ~90% — no sooner, so you're not re-drilling things you already own.
Ranked weakest-first by a confusion-aware score. Tap “I'm not sure” on any term and it jumps the queue — without denting the rating that reflects what you can actually do.
Concepts are wired into a prerequisite graph, so you meet attention only after you've met softmax. The frontier is whatever you've just unlocked — never random.
Longest-unseen concepts, mixed with a bandit that probes the topics the model is least sure about you. It keeps the whole map warm instead of letting corners go stale.
Two halves of the same idea — your knowledge has a rating, and that rating is earned.
An ELO-style rating on the same 800–2400 scale chess uses. Every card has one. You have one per domain. When you answer, both ratings update — the way a chess rating moves after a game. Cards that prove harder than predicted climb; cards that prove easier sink.
Ranked, timed, head-to-head matches. You're paired with an opponent at your EigenScore, served the same cards, and scored on accuracy and speed. Wins move your rating up, losses move it down. The same engine that grades cards grades players.
Every track has an allowlist of authoritative domains. Cards on time-sensitive topics — LLM tooling, library APIs — are re-validated on a weekly cron. Here's the starting list.
Eigencards is organized into Fields of Study. Data Science & AI is live now — eight domains, hand-mapped into prerequisite graphs so the app teaches in order, not at random. All eight are open from day one; the engine calibrates where you stand and starts you at the right depth. New subjects come next; the engine doesn't care what you're mastering.
Learn for free — every domain, the full difficulty ladder, a daily deck of cards. Out of cards? Watch a short clip for more. Or go Pro and lift the limit entirely — unlimited cards, no ads, and the Arena.
If we don't know yet, we'll say so.
