Personal project2026–ongoing
Quaderno di Statistica: tutoring with AI-built lessons
An interactive notebook to prepare a first-year statistics exam, built lesson by lesson around one student's results.
- My role
- Tutor; designed the method and the lessons, built with Claude
- Data
- The student's results after each lesson (kept private)
The question
A first-year university student had to pass a basic statistics exam: confidence intervals, hypothesis tests, two-group comparisons, regression. Textbooks explain the formulas; what she lacked was knowing which formula a question asks for, and why. Could short lessons, adapted after every session to what she actually got wrong, get her there faster than another pass through the slides?
What I did
- Set up the course: an index of the course materials (which stay private, they belong to the teachers), then a diagnostic “lesson 0” over the whole syllabus.
- Planned and taught the lessons in person, 60–90 minutes each.
- After each lesson and each homework set, read the results and decided what to change next.
- Built the notebook with Claude (Anthropic): lessons, two to four homework variants per lesson, eight exam simulations, a printable formula sheet and an R command sheet. Every number was checked in R.
The method
One cycle per topic: lesson in person → results → 2–4 homework variants → results → next lesson.
- Every page records the answers: right and wrong, confident-but-wrong, self-assessment of the paper exercises, doubts, time. The student sends the summary to the tutor.
- After each result, a short learning record: what was observed, how I read it, what changes. These stay private.
- Recurring toy stories carry the concepts: three coins (fair, “heads”, “tails”) for hypotheses and tails, a race with a mystery drug for tests on means, two teams of athletes for comparing groups.
- Interactive tools to see the ideas: the normal curve and its quantiles, the test↔interval window, a hundred repeated experiments, the knobs of an interval’s width.
- Short rules to apply every time (“observed minus hypothesised”, “an interval is always two-sided”), and spaced repetition of the recurring mistakes.
What I learned
- The diagnosis paid off: the theory was there, the bottleneck was choosing the distribution and the quantile, not the arithmetic.
- Building the confidence interval as “a test in reverse” made the step from tests to intervals feel natural.
- The persistent errors were conceptual and recurring (θ₀ or θ̂ in the standard error, α/2, the “95%” attributed to the parameter), and short rules plus repetition worked better than re-explaining.
- An AI assistant makes this kind of tailoring affordable: new variants with fresh numbers and complete solutions after every session, which by hand would take days.