Start class
Start with the class you already have.
Use the microphone in a live room, add tab audio for YouTube or Coursera, or watch a video link inside Learnica. Every path opens the same live workspace.
Used by 50+ students in 6+ countries and counting
See what Learnica handles, and what it leaves with you.
The story begins with a familiar classroom gap, then follows one class through live capture, notes, recall, and the learning science behind each choice.
Learnica keeps upwhile you
Learnica captures the class and keeps the page moving, then asks you to predict, choose, or recall at a few useful moments. You get the ease of Otter or Granola while more of the thinking remains yours.
Before class
Give the lesson somewhere to land.
Learnica asks for a quick prediction before class begins. The question is optional, but answering it gives the coming explanation something to connect to.
In class
Let the recording keep up.
Live captions, translation, and a complete transcript catch the words you miss. The page stays as flexible as GoodNotes or Notability without making you chase every sentence.
While notes generate
Turn the wait into a memory check.
Every AI note taker needs a moment to turn a recording into notes. Learnica uses that same pause to ask what you can recall before the finished page opens.
After class
See what stayed with you.
The concept mirror compares your recall with what the class covered and points to what needs another look. Otter, Granola, and Notion AI show what was said, while Learnica also shows what reached your own words.
Coming next
A tutor that starts with a hint.
The next Learnica tutor will catch you up or unpack a difficult point using the actual class materials. It will add more only when your own reasoning needs it.
Keep the effort that changes what a student can do later.
The same task can help one learner and overwhelm another. Learnica treats support as a hypothesis, then looks past the finished page to unaided recall, transfer, and calibration.
Useful difficulty is not a fixed amount of friction.
It depends on what the learner already knows, what the task asks, and what must still be possible when the tool is gone.
productive struggle
Protect the work that creates meaning.
Prediction, selection, retrieval, and self-explanation ask the learner to construct or recover an idea, so Learnica keeps those moves in the flow.
- Predict before an explanation
- Choose and rephrase the main idea
- Recall before reopening the notes
- Explain why an answer works
wasteful struggle
Take on the work that spends attention without teaching.
Racing to transcribe, recovering lost audio, and decoding a second language can consume capacity without strengthening the target idea, so Learnica carries that load.
- Race to transcribe every sentence
- Decode a second language in real time
- Recover audio lost to a short interruption
- Copy the same material again after class
These are design judgments, not permanent labels. The assistance dilemma is deciding when support removes waste and when it removes the thinking that produces learning. Prior knowledge, timing, scaffolding, fading, and expertise reversal can change that answer. Learnica tests each change against delayed unaided recall, transfer, metacognitive calibration, and the learner's experience instead of treating a cleaner artifact as proof of learning.
Four moments turn that learning contract into product behavior.
Each one has a plain interaction at the surface and a testable mechanism underneath.
Before class
Try an answer before you see one.
A quick guess gives the coming explanation somewhere to land, even when the guess is wrong.
Pretesting asks learners to attempt retrieval before content. It improves later memory with an effect of d=1.1 at 95% error (Richland, Kornell and Kao 2009).
While taking notes
Give the page shape without filling it in.
Headings and prompts remove the fear of a blank page while leaving the meaning for the student to build.
Encoding improves when learners select, organize, integrate, and self-explain. In cognitive load terms, the skeleton removes transcription burden while protecting germane load. Guided outlines outperform complete notes at d≈0.55. A scaffold should fade as knowledge grows because of expertise reversal (Larwin and Larwin 2013, Kalyuga 2007).
Before notes open
Ask memory before opening the notes.
Bringing an idea back strengthens it in a way that reading the same idea again does not.
Retrieval practice produces a testing effect of g=0.50 across 61 studies (Rowland 2014). Learnica places the attempt inside the generation wait, so it does not add another waiting step.
During review
Compare memory with evidence.
A clear mirror helps students separate an idea they understood from one they only remember seeing.
Metacognitive calibration is part of self-regulated learning and is weak without external evidence. Unaided judgment tracks actual learning at roughly γ≈0.27 (Koriat and Bjork 2005), so Learnica scores the student's own recall against a concept list committed before generation.
The artifact is not the outcome.
A finished note can look better while learning gets worse. The product therefore separates performance with help from what remains when the help is gone.
- With support
- Can the learner stay oriented and make a useful attempt without spending the class racing to copy?
- Delayed recall
- What can the learner bring back after time has passed and the finished notes are closed?
- Transfer
- Can the learner use the idea in a new question rather than repeat the wording that appeared in class?
- Calibration
- Does confidence track what the learner can actually explain, and does the concept mirror improve that judgment?
What happens when the AI does more?
Two studies make the same tradeoff visible. A polished result can improve while the learner becomes less able to work without the tool.
Figure 01Post-test score
Moderate help beat complete automation.
Figure 02Relative change in performance
Hints protected the exam while improving practice.
ConditionPracticeExam
Standard GPT+48−17
Hints-only tutor+1270
Four ways to leave class with notes.
The page can look complete in all four. The difference is who had to keep up and whether the learning process is visible afterward.
Pen and paper
- Who keeps up
- You
- What the page keeps
- Only what you write
- When thinking happens
- Throughout class
- What review shows
- Your notes
GoodNotes and Notability
- Who keeps up
- You
- What the page keeps
- Your writing and the media you add
- When thinking happens
- Throughout class
- What review shows
- Your notebook and linked media
Otter, Granola, and Notion AI
- Who keeps up
- The AI
- What the page keeps
- A transcript or polished summary
- When thinking happens
- Whenever you choose
- What review shows
- The record and summary
Complete capture with your thinking
- Who keeps up
- The AI
- What the page keeps
- The class and your own attempts
- When thinking happens
- At small moments in the flow
- What review shows
- The record and what stayed
A classroom AI system has to earn every answer.
A fluent response is the easy part. The system has to survive the room, retrieve the right evidence, release only the help the learning moment allows, and show where it came from. Learnica separates capture, evidence, policy, generation, and evaluation so one polished answer cannot hide a broken source or a missing learning step.
See one question move from thought to answer.
The live product captures the class, shapes notes from class materials, and compares recall. The next build adds a tool-using tutor that releases help in stages.
Start with a prediction.
Learnica asks the learner to commit to one idea before the lecture. The answer stays locked, while capture and translation remain available.
Why can adding more context make an answer worse?
✓Before class2 of 3
✓Preview the question
Write one prediction
Join the lecture
More context can bury the passages that matter, so the model may follow the wrong evidence.
Your own words
Today, help waits for a thought of your own. The planned engagement check will look for meaningful language rather than count characters.
The machine keeps pace while the learner decides what matters.
Complete captions and translation follow the room while the note stays short enough to belong to the student.
Live transcriptRunning now
A capable model can still answer badly when the context is weak.
Retrieval first finds relevant passages, then adds them to the prompt as evidence.
More context is not automatically better. Retrieval and generation need separate evaluation.
StreamStateRecord
MicrophoneLiveLocal copy
TranscriptShapingComplete sentences
TranslationFollowingBounded window
Retrieval before generation
RAG lets a model look things up before answering, but the retrieved context still has to be relevant.
Quiet assistanceNext build
Matching your note to the lecture sources
Smallest useful hintNext build
You may want to add why context precision matters.
Keep the learner's words in view.
Learnica keeps the learner's shorthand visible beside additions grounded in class sources. The next provenance layer will make the source of each addition directly inspectable.
+3-1
1-
RAG lets a model look things up.1+
Retrieval finds passages before generation.2+
The passages become evidence in the prompt. [1][2]3+
Retrieval and generation need separate evaluation. [3]A longer context can still fail when the useful evidence is difficult to find.123
Lecture transcript18:42
Course deckSlides 12 and 28
Diagram inside the noteNext build
QuestionRetrieveAnswer
Generating a retrieval map
Use the class before using the web.
The tutor retrieves and verifies course evidence first, then chooses the smallest useful response.
Why can more context make a RAG answer worse, and how would I test it?
Course sources first
Evidence trace
Matched the question to the transcript at 18:42 and the RAG diagram on slide 12.
Retrieved the component and end-to-end evaluation framework from slide 28.
Course evidence was sufficient, so web search did not run.
AnswerGrounded in class
A longer context can contain more evidence and still hurt an answer when relevant passages are harder to identify. Test retrieval and generation separately, then evaluate the final response.12
When the question asks what changedNext build
Searching recent RAG evaluations
- ACL Anthology research index
- arXiv research index
- OpenReview research index
Next build · Python
def retrieve(query, course, k=4):
hits = lexical(query, course, k=20)
if low_confidence(hits):
dense = semantic(query, course, k=20)
hits = rrf(hits, dense)
return rerank(query, hits)[:k]✓Review plan1 of 3
✓Explain retrieval without notes
Compare recall with the class record
Build the compact implementation
ConceptYour recallClass record
Retrieval firstCapturedSupported
Context precisionReviewSupported
End-to-end evaluationCapturedSupported
The system proves itself when the room stops behaving like a demo.
One trace follows a live class through an audio failure. The other follows a request for help through the planned tutoring policy.
Running today
A headset dies in the middle of a sentence.
A student may switch devices, lose a headset, or pass through a short network interruption while the lecturer keeps speaking.
Source changesA headset or tab stream disappears.
Local audio holdsUnsent chunks remain in IndexedDB.
Capture resumesThe same session continues after the device changes.
Text rejoinsSentence boundaries and translation catch up.
What the student seesThe class continues through a short interruption without discarding the buffered passage.
Building next
A learner asks for help before making an attempt.
The policy chooses a learning move before any answer is drafted, so orchestration does not quietly become answer generation.
Attempt checkedThe learner's own words come first.
Move chosenRecall, question, or hint fits the moment.
Evidence gatheredThe transcript, deck, and course memory stay in scope.
Claim verifiedSources and release policy must agree, or release stops.
What the student seesThe learner gets the next useful move, not a finished solution.
A messy room still becomes a readable record.
Learnica keeps streaming when an audio source changes or the network drops briefly, then returns captions as complete sentences with translation moving beside the lecture.
A stable capture graph resumes after device changes. Audio chunks persist in IndexedDB before upload, while sentence-boundary processing and bounded translation windows keep multi-hour sessions responsive.
The unlock will recognize a real attempt.
The current gate asks for a thought of the student's own before AI help becomes available. The next version will check for meaningful engagement rather than count characters.
Trivial keystrokes are not the design intent. The planned check asks whether the entry is in the student's own words without turning the gate into surveillance or a writing test.
The class remains the boundary.
The deck, transcript, and concept list define what generated notes may cover. The student's own words remain distinct from the generated record, and the source material stays available for review.
Retrieval-augmented generation is scoped to the current class before synthesis. The next provenance layer will attach stable source identifiers to generated additions. The concept mirror already compares recall against a target committed before generation rather than a list invented after the answer is known.
Use an agent only when the next learning move is genuinely uncertain.
Capture, gates, and release remain deterministic. Planning begins only when the learner's attempt, the available evidence, and the help policy leave more than one valid next move.
A typed state machine owns the session. Planner, retriever, coach, and verifier receive bounded roles and schemas. Course uploads and web results remain untrusted evidence, so retrieved instructions cannot redefine tool permissions. Tools run with least privilege. The interface exposes sources, tool outcomes, and release decisions rather than private chain of thought, while prompt injection tests remain part of the release evaluation.
A missing condition never becomes a convenient guess.
Each student receives a server-held condition for each lecture. If that state is missing, the intervention stays closed rather than assigning a condition on the client.
The experiment path fails closed. Each effort feature reports telemetry against a declared success signal, and recall judgments require blind human validation at κ ≥ 0.61 with a reported confusion matrix before automated scores count as evidence.
Deeper retrieval has to earn its delay and cost.
An exact course term should be found quickly. Semantic retrieval, reranking, graph traversal, or the open web enter only when the question needs them, and remembered claims stay visible to the learner.
Lexical retrieval comes first. Reciprocal-rank fusion and cross-encoder reranking enter when ambiguity justifies them, while graph traversal is reserved for questions that depend on relationships across classes. Web search runs only when course evidence is insufficient or freshness matters. Retrieval precision and recall, generation faithfulness and relevance, latency, cost, and end-to-end learning are evaluated separately. Course memory stays scoped to the learner and course, and remains inspectable, correctable, and deletable.
I kept thinking about the questions that came back blank.
Before Stanford, I taught more than 1,000 undergraduates across seven computer science courses. As generative AI coding assistants became part of everyday coursework, I sometimes saw elegant take-home code followed by silence when an in-person final asked a simple question we had worked through many times. There was no need to assume cheating, because the work could be genuinely theirs while too much of the thinking still happened somewhere else.
At Stanford, I began treating that classroom question as both a learning-science problem and a product problem. I brought years of building full-stack AI systems at Apple, TikTok, Alibaba, AMD, and early-stage startups to a tool that now serves about 100 students. Learnica keeps the complete record of a class, follows the room with captions and translation, and creates a few well-timed opportunities to predict, choose, recall, and explain. The aim is to make a difficult class easier to enter, wherever a learner begins, without letting a polished artifact stand in for understanding.
Education becomes personal when a teacher notices the hesitation behind a polished page, remembers an earlier struggle, and stays long enough to help the idea land. Learnica can carry the record, keep the evidence close, and make more room for those moments. What it cannot do is care whether the idea finally lands, because that responsibility and that relationship must remain human.
Zikun "Zayden" Zhu