Inspiration
At school, revision is planned for you. Teachers set the homework, the tests come every week, and rereading your notes the night before usually gets you through. Then you arrive at university and that structure disappears. You have a 300-page reading list, one exam at the end of term, and the methods that worked at school quietly stop working. Most of us found this out the hard way, in our first set of exams.
It is harder still for students whose focus comes in bursts, including many students with ADHD. Every study app hands you the same fixed 25-minute timer, and it either cuts you off when you are finally in the zone or keeps you staring at a page long after your attention has gone.
Learning science already knows which techniques tend to work: practice testing and spaced practice beat rereading and highlighting (Dunlosky et al., 2013). What it can't tell you is which one works for you, on this subject. We wanted an app that finds out.
What it does
- Tries three methods on your own notes. Active Recall, Practice Test and Feynman (explaining it to "Sam", an AI student), each scored on the same 5-question quiz so the comparison is fair.
- Learns which one teaches you the most per minute, and keeps testing the others now and then in case that changes.
- Watches your pace, not a clock. Answer timing is compared with your own baseline. When your answers get slow or erratic, Attune offers an easier question, a different format or a break.
- Works out why you stalled. Tired, away, plateaued or already mastered? Each has a different fix. On a plateau, Attune runs a one-week trial of the runner-up method and keeps it only if your scores actually rise.
How we built it
Three of us built it in one day: a FastAPI + SQLite backend, a React + TypeScript frontend, and an engine of pure Python functions with no network or database access, so every decision can be tested in milliseconds.
Choosing a method
Choosing a method is a multi-armed bandit. Each method has a Beta distribution, seeded from the learning-science priors above. After each session the reward
$$ r = 0.60\,\text{retention} + 0.25\,\text{efficiency} + 0.15\,\text{felt} $$
updates it with \( \alpha \leftarrow \alpha + r \) and \( \beta \leftarrow \beta + (1 - r) \), and Thompson sampling picks the next method, with at least 15% of sessions still exploring.
Reading attention
Attention is read from answer-time variability: a trimmed coefficient of variation over the last 8 answers, compared with your own baseline. It ignores one slow answer but catches two or more, and the state only changes when two readings in a row agree, so it doesn't flicker.
Diagnosing a stall
Diagnosis uses a Theil–Sen slope and a CUSUM check on quiz scores, alongside engagement and focus signals, to tell tiredness apart from a real plateau.
The AI never makes the decisions
Language models hallucinate, so every plan change comes from transparent rules over your own data. The LLM only marks short answers and explains the decision in plain words. If its explanation goes off-script or mentions anything medical, we discard it and show a template instead.
Challenges we ran into
- Our attention detector was too robust. We first used median-based statistics so one slow answer couldn't trigger a break. They also ignored a genuinely erratic burst until half the window was erratic. Switching to a trimmed spread fixed it, and the demo burst is now caught two answers sooner.
- The agent tested the wrong method. In an end-to-end run, the weekly reflection "trialled" the method our student was already using, because the current plan happened to be an exploration pick. We now diagnose the method actually studied, and lock that story in with tests across several random seeds.
- Roles changed mid-build. We swapped responsibilities an hour in and kept the API contract frozen, so all three parts still fitted together.
What we learned
- Every threshold is a trade-off. Make it robust to one thing and it goes blind to another.
- Test the demo path end to end, not just the units. Both of our biggest bugs passed every unit test.
- An LLM is a great narrator and a risky decision-maker. Keeping it out of the decisions made the system easier to trust and to test.
Honest limits
- The multi-week history in the demo is seeded for a simulated student, and labelled "Demo data" in the app.
- Attune is attention-aware, not a medical tool. It doesn't detect, diagnose or treat ADHD or anything else.
- Our thresholds are starting values tuned on simulated students, and readiness hasn't been checked against real exam results.
What's next
- A delayed quiz to measure long-term retention
- Validation against real exam results
- A live voice tutor for Feynman mode