The Story
This app exists to solve one problem: “I want to easily understand and memorize AI engineering concepts.” Everything on this site — every feature, down to the choice of database — is derived from that sentence.
The problem
Learning AI engineering the obvious way fails twice. Ask a chatbot to explain a concept and it answers confidently from memory — sometimes wrong, and a confident wrong answer is worse than none. And even when the explanation is right, you read it once and it is gone in a week. Two failure modes, one sentence:
AI makes things up, and I forget what I learn.
How the AI fixes it — four words
The fix is one loop with four steps. The first two make what you read trustworthy; the last two make it stick.
The AI searches the real source material first — it never answers from memory.
Sources are chunked and embedded into a vector index; a hybrid retriever (QueryFusionRetriever — dense meaning-search fused with BM25 keywords) pulls the passages that actually match your question.
It writes only from what it found, then a second step checks every claim is backed — or it isn't shown.
CitationQueryEngine puts a [n] source on each sentence; FaithfulnessEvaluator scores the answer against the retrieved context. Unsupported → regenerated once, stricter; still failing → dropped, not published.
It predicts the moment you are about to forget each fact.
Every review updates an FSRS-6 memory model — difficulty, stability, retrievability — for you and that specific card, tracing your personal forgetting curve.
It shows you the card right then — not too early, not too late.
The scheduler surfaces each card at the predicted edge of forgetting, interleaving material the research says benefits from mixing. Your grades feed the next prediction.
One line
It looks it up, proves it’s true, times your forgetting, and repeats it.
One loop, not two features
Understand and memorize are not separate features — they are one instrumented loop:
1. Ground. Real sources are retrieved and an explanation is generated only from them (LOOK).
2. Gate. A faithfulness check verifies the explanation against those sources before it ships (PROVE).
3. Learn. You read the vetted explanation — every sentence cited.
4. Recall. The scheduler predicts when the memory will slip and quizzes you right then (TIME + REPEAT).
5. Close the loop. Your grades update a mastery model that decides which concepts get re-explained next — recall feeds understanding.
That is also why the storage is what it is: a vector store because understanding needs semantic retrieval, an edge database because memorizing needs durable per-user schedules, object storage because audio explanations are large. Each store is the cheapest thing that satisfies one of the two verbs.
How AI-native is each step?
Honest answer: every step is a model decision wrapped in a deterministic envelope — the model proposes, the math clamps. That pairing, repeated four times, is the design.
A model decides how to search for each question — meaning-search, keywords, or a multi-step retrieval — and a cross-encoder re-scores the finalists. The envelope: a classic keyword lane runs alongside for exact terms, and if the model is unsure the system falls back to a fixed strategy. AI picks the approach; deterministic rules guarantee it never comes back empty.
AI produces and AI verifies: a model writes the explanation from retrieved sources, then independent judge passes check every claim against those sources. The envelope is code — each citation must resolve to a real source, unsupported claims are rewritten once, then dropped. Nothing ships unproven.
A model now tailors the review schedule to you — estimating how hard a new idea is, adapting to interference between similar cards. But every result is clamped to a tight band around the proven forgetting math, so a model can tune the timing and never invent an unsafe date. AI proposes; the math sets the guardrails.
A tutor agent composes your review session — choosing and ordering what to show, and writing fresh, self-verified cards. The envelope: it can only pick from what is actually due, enforced in code, and it never rewrites a card you have already rehearsed. AI shapes the session inside a boundary it cannot cross.
The rule behind all four: AI where judgment is needed, math where correctness is needed. Relevance and groundedness are judgment calls, so models decide them. A review schedule is a contract, so the math clamps what the model is allowed to change.
Recall check
This page ends the way the app works — try to answer before revealing. No peeking.
What two problems does this app exist to solve?
Show answer
AI chatbots make things up, and you forget what you learn — trust and retention.
From the research: Retrieval practice (the testing effect) — Roediger & Karpicke — Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention (2006)
Say the four words of the fix, in order.
Show answer
LOOK · PROVE · TIME · REPEAT.
From the research: Retrieval practice (the testing effect) — Roediger & Karpicke — Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention (2006)
Which two words stop the AI making things up, and which two stop you forgetting?
Show answer
LOOK + PROVE make it trustworthy (retrieve, cite, faithfulness-check); TIME + REPEAT make it stick (predict forgetting, schedule the review).
From the research: Retrieval practice (the testing effect) — Roediger & Karpicke — Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention (2006)
What single rule lets every step use AI without risking wrong facts or unsafe timing?
Show answer
AI proposes, math clamps. A model makes each judgment, but a deterministic envelope — citations that must resolve, a schedule clamped to the proven forgetting curve, a due-set the composer cannot exceed — guarantees it can never ship an unsupported fact or an unsafe review date.
From the research: Retrieval practice (the testing effect) — Roediger & Karpicke — Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention (2006)