An iPad tutor, built for my kids
A difficult problem can be the beginning of a good conversation. トイ helps children work through it, find the idea worth keeping, and meet that idea again later.
Japanese tutoring · Native iPad app · Multiagent backend
Photo, PDF, or typed problemRoom to draw and explainA fresh problem later
The learning goal
Finishing a problem together is useful. The question is whether the child can recognise the same idea and use it independently on a different problem, days later.
That goal shapes the whole loop: the child produces the reasoning; the tutor gives enough support to keep moving; the lesson is put into words; a fresh problem returns later with less help.
Why assisted performance and learning need separate checks ↗
Made for the iPad
The question stays in view while the child talks, draws, and asks for help. A question of their own can start a different kind of conversation.
Fresh captures of the current app in the iOS simulator, displaying anonymized archived sessions. Names are replaced and source photos are omitted; the conversations and generated diagrams are from the records.
From the homework table
A few moments from actual sessions: getting stuck, changing an idea, and returning to it. English translations are paired with the Japanese originals. Names are removed; omissions are marked.
A reading question asks what a speaker is doing when they bring several people’s ideas together. The child understands the passage but is stuck finding the wording for a short answer.
01
FIRST SESSION · THREE CONSECUTIVE MESSAGES
The tutor directs attention to a phrase in the passage. The child supplies the term. It is help with expressing an idea they have already understood.
02
LATER IN THE SAME SESSION · INTERVENING TURNS OMITTED
The saved lesson describes where to look and what to look for. It can be used beyond the wording of this particular passage.
03
TEN DAYS LATER · A NEW PASSAGE · THREE CONSECUTIVE MESSAGES
The review is linked to the same saved strategy. The child supplies the answer and identifies the textual clue. This is a closely related problem with the same cue word and answer pattern; it does not establish broad transfer.
The child is counting when clock hands overlap, starting at 2 a.m. They have guessed that the ninth overlap falls in the eleven-o’clock hour.
Four consecutive messages show the tutor moving from a question to a concrete example of how to count. The child completes that step; calculating the exact time comes later.
While correcting a microscope question, the child has said that brightness should be adjusted before adding the slide. Asked why, they say it was just a feeling. The tutor checks the reasoning.
Later, after further explanation and the corrected answer · intervening turns omitted
The child’s first correct choice did not reveal their belief about the specimen. A question tied to the actual experimental conditions made that belief visible and gave them a reason to reconsider it.
This is a question the child brings to the curiosity conversation, without a worksheet.
Further explanation, a drawing and the child’s thanks omitted
The tutor gives information directly and closes with the child’s own short account of what they found out. Curiosity has a different rhythm from working through an assigned problem.
An archived moving-points session also records a clear failure. These are four consecutive messages; two tutor replies are shortened as marked.
No animation was attached to that reply. The current implementation has a dedicated moving-figure renderer and controls; the historical exchange shows why claiming a capability is not enough. The conversations expose failures as well as useful teaching.
Research into practice
The useful detail is how a finding changes the next question, the amount of help, or what happens after a session.
Research: explaining, predicting and generating a representation can engage learners more deeply than following an explanation.
In トイ: the child retells the problem, attempts a step, and explains the idea behind it. Routine calculations can move quickly; questions about why are saved for the steps that carry the concept.
Research: discovery needs guidance. Missing knowledge and repeated failure call for information; the right support depends on prior knowledge.
In トイ: the tutor aims for one short question at a time. The server tracks unsuccessful attempts and specifies stronger help; after three on a step, it calls for an explanation or a parallel example. Younger learners start with more guidance; familiar strategies start with less.
Research: visual representations need to be taught. A useful diagram makes quantities and relationships easier to reason about.
In トイ: a diagram can come with a small example or an outline to complete. The tutor is instructed to invite the child’s attempt before showing its own drawing. Moving figures ask for a prediction before each step, with advance, replay and back controls.
Research: reflection is more useful when it identifies an accurate, transferable idea. Japanese work on 教訓帰納 studies this process of drawing a lesson from solving or getting stuck.
In トイ: the child states what to notice and what to do next time. The tutor helps refine that into a strategy card. A second method is introduced after the first is understood, with an explicit comparison.
Research: spacing practice and mixing problem types can improve later performance. Mixed practice makes choosing a method part of the work.
In トイ: reviews use fresh problems for the same strategy, with its name hidden at the start. Due reviews are offered before new work, and the amount of help recorded affects scheduling. FSRS supplies an approximate schedule; its fit to children’s problem solving is still an open question.
Research: motivation depends in part on feeling capable, having a say, and feeling supported. Reward effects depend on their design and context.
In トイ: feedback names a useful action. XP recognises trying, drawing, noticing an error and explaining a strategy, including unfinished work. Study days accumulate, hints are free, and the child can stop or return later.
Beyond the worksheet
“Why does a rainbow appear?” gets a direct explanation at the child’s level and room to follow their curiosity. It has its own conversation mode, without the worksheet’s answer-withholding rules or scheduled reviews.
The tutor asks what the child wrote, what the key says, and how sure they were. Together they examine the reasoning and record a concrete check to use next time. A later variant lets them revisit the idea.
A weekly letter draws on recorded work. The child chooses or writes an “if I get stuck, then I’ll…” plan. The parent sees strategies, sticking points and help needed, with something specific to ask about. The child can see that report too.
Something to carry into next week
The weekly letter looks back at recorded work. The child chooses or writes one response to getting stuck, and the parent gets something specific to ask about.
View the screenshot ↗
The research behind it
Ten research threads shape the teaching decisions. Open a topic to see the finding, what it changes here, and where the evidence stops.
The literature includes classroom experiments, research syntheses, small case studies and theory. Their ages, subjects and teaching settings differ from this app’s.
The exact help ladder, review intervals, rewards and weekly letter are implementations informed by that work. They have not been validated together as a teaching method.
トイ has not demonstrated improved learning in a controlled study. Testing whether a tutor follows its rules does not establish retention, transfer or better exam results.
A high-school mathematics experiment found that an answer-giving AI could improve practice performance while harming later performance without it. A version using teaching safeguards largely mitigated that harm. This motivates a separate check of independent work. Bastani et al., 2025 · PNAS.
A university physics trial reported gains with a purpose-built tutor. Its authors also found that prompts alone did not reliably maintain the teaching sequence, so they structured it in the platform. This informed the decision to keep parts of トイ’s teaching logic in code. Kestin et al., 2025 · Scientific Reports.
Both tested specific systems in specific settings. Neither establishes an effect for this app or for younger children studying at home.
The ICAP framework distinguishes receiving or manipulating information from generating an explanation and building on another person’s reasoning. It informs the request for a reason, a prediction or a drawing, rather than a succession of yes/no answers. Chi & Wylie, 2014 · Educational Psychologist.
A meta-analysis distinguished unassisted discovery from discovery supported by feedback, examples and scaffolding. The distinction matters here: the tutor can explain a missing step and then hand the work back. Alfieri et al., 2011 · Journal of Educational Psychology.
These support guided learning in broad terms. They do not validate this six-level ladder, its thresholds, or an LLM’s ability to diagnose frustration.
In two experiments with 11–12-year-olds, mathematics practice spaced across weeks produced better month-later test performance and more accurate confidence judgments than practice in one session. Emeny, Hartwig & Rohrer, 2021 · Applied Cognitive Psychology.
A classroom trial in seventh-grade mathematics supported interleaved practice, in which successive problems require different strategies. This motivates reviews that make recognising the method part of the task. Rohrer et al., 2020 · Journal of Educational Psychology.
The app schedules due strategies and hides their names; it does not enforce an alternating sequence of problem types. Fresh variants, reduced help and FSRS scheduling are its adaptation. An optimal interval or daily dose for these whole-problem reviews has not been established.
The IES mathematics guide recommends teaching visual representations, monitoring the solving process and making mathematical concepts explicit. It also cautions that multiple strategies need careful introduction and comparison. These inform the path from the situation to a diagram to an expression. Woodward et al., 2012 · IES practice guide.
Japanese research on 教訓帰納 examines extracting an accurate, transferable lesson from an error. In one junior-high study, lesson quality predicted later performance on some of the items studied. That informs the strategy card’s focus on a specific method or check. Shiba, 2023 · Japanese Journal of Educational Psychology.
A review of science education research found that texts which explicitly address a misconception can support conceptual change better than ordinary exposition. トイ’s science instructions ask the tutor to elicit a prediction, explain why a mistaken idea fails, and give the accepted account. Tippett, 2010 · International Journal of Science and Mathematics Education.
The Japanese finding is correlational; the science review concerns texts, not AI dialogue. A reflection or one corrected answer does not establish lasting conceptual change.
Self-determination theory connects motivation with autonomy, competence and relatedness. Its educational review distinguishes supportive structure from pressure and control. It informs the child’s choices, specific feedback and the ability to ask for help or stop. Ryan & Deci, 2020 · Contemporary Educational Psychology.
トイ rewards ways of working, keeps hints free, avoids competitive rankings, and lets study days accumulate without a broken streak. A parent sees what the child tried and where help was needed.
Those are design choices. The research does not establish that this XP system improves intrinsic motivation or learning; returning to an app and learning from it are separate outcomes.
Reciprocal teaching organises reading dialogue around summarising, questioning, clarifying and predicting, with adult modelling and guidance. These moves inform トイ’s reading sessions. Palincsar & Brown, 1984 · Cognition and Instruction.
The broader evidence review by Dunlosky and colleagues rates practice testing and distributed practice highly. For social studies, the app supplies missing facts, asks about their connections, and later prompts recall. Dunlosky et al., 2013 · Psychological Science in the Public Interest.
The source settings differ from a Japanese child reading with an AI. The subject-specific prompts are adaptations, not a tested transfer of the original interventions.
In experiments on understanding a mechanical system, asking learners to predict motion from a static diagram helped comprehension; showing animation alone did not outperform static diagrams. This informed the prediction before each moving step. Hegarty, Kriz & Cate, 2003 · Cognition and Instruction.
Experiments with narrated animation also found transfer benefits from letting learners advance through segments. トイ applies that idea to short steps in a moving figure. Mayer & Chandler, 2001 · Journal of Educational Psychology.
These did not test this app’s diagrams or its learners. The drawing still needs to be accurate and appropriate to the problem.
Studies of preschool conversations found that children often pursued follow-up questions after an explanation and repeated their original question after a non-explanation. That supports treating a spontaneous “why?” as a request for information. Frazier, Gelman & Wellman, 2009 · Child Development.
In トイ, curiosity has its own mode: explain plainly, follow the child’s interest, and close with what they found out. Homework questions still use the guided problem-solving path.
The cited work concerns preschoolers and human adults. It motivates this design choice without establishing effects for AI conversations with older children.
A small randomised study taught fifth-graders to contrast a desired outcome with obstacles and form if–then plans. It reported benefits for school outcomes. This is one influence on letting the child choose a concrete response to getting stuck. Duckworth et al., 2013 · Social Psychological and Personality Science.
The weekly letter draws on recorded work and the child’s strategy cards. A chosen plan can be brought back once when they next get stuck. The parent gets a concrete conversation starter.
The study tested mental contrasting together with planning, not an automated weekly letter or an if–then sentence alone. The app’s weekly format remains a design hypothesis.
A field experiment with seventh- and eighth-graders learning fractions compared reflecting on the reasoning behind their own errors with reflecting on the corresponding correct solutions. The former group showed greater procedural knowledge, with a conceptual advantage at follow-up. Heemsoth & Heinze, 2016 · Journal of Experimental Education.
トイ’s correction flow records the child’s answer, the key and their confidence, then works towards a specific error pattern and a check for next time.
This evidence concerns a particular mathematics intervention. The full correction workflow across subjects, and the choice of support from a confidence rating, have not been validated here.
Can the child recognise and solve a fresh problem later, with less help, and explain why the method fits?
Under the hood
Independent agents read and solve the problem. Once the reading is settled, they draft teaching approaches that are combined into a plan.
When readings disagree, a referee checks them against the source. An unresolved reading is marked for a parent. The conversation can proceed from the checked reading while the teaching plans are prepared.
Code keeps the session state, calculates help levels, checks replies and schedules reviews. Teaching prompts guide the diagnosis, questions and explanations. A flagged reply gets one rewrite; a second reply may still break a rule. Even when agents agree, a reading or solution can be wrong.
SwiftUI iPad app / Python + FastAPI / SQLite / FSRS review scheduling
A family project
I built this while my kids were preparing for exams. They were turning to an AI assistant for solutions; I wanted to give them somewhere to work through the part they didn’t understand.
It started with math and grew from there. The project now brings together research on tutoring dialogue, memory, drawing, motivation and reflection. Watching my kids use it keeps raising the next design question.
Keep the conversation going
I shared the project on Hacker News. The conversation continues there.
