Clear View

A project that helps students view the tasks at hand. It can help identify whether the student may pass or fail the course as well as show resources to help guide them through the year.

Comment

inspiration

as students ourselves, we have been there. students keep deadlines in one place, notes in another and grades in a third, and usually find out they are in trouble when it is too late to fix. we wanted one place that answers three questions: what do i need to do, what do i need to study, and where do i need support.

what it does

  • one real student, everywhere: the dashboard follows a student from the wolfhacks dataset. switch the student id and every page updates.
  • overview: a calendar where you add, tick off and remove deadlines and exams, with "due soon" counted from your list.
  • my progress: twelve numbers for the semester (midterm, attendance, on-time rate, study hours and more), each next to the average of the other students in the same course.
  • notes: upload a pdf, image or text file and get a summary, key points and a clickable practice quiz, generated by google gemini.
  • support check: a decision tree estimates whether a course needs "extra support", "some support" or is "on track", explains why in plain language, and gives a next step for each reason. change a number and check again to see what would move the result.
  • spookie's pond: a rubber duck in a witch hat for talking through what's stuck in a class. spookie knows what the dashboard shows for you, asks one question at a time, and ends with one small step.
  • resources: links to campus academic support, so a next step always has somewhere to go.
  • how the model works: recall, baseline comparison, the overfitting chart, the tree itself and its limits, on one page.

how we built it

  • model: a scikit-learn decision tree trained on the 800-row wolfhacks synthetic dataset. we held out 20% (160 rows) before doing anything else, and chose the tree depth and the decision cutoff with 5-fold cross-validation on the training rows only.
  • data cleaning: three percent columns converted to fractions. blanks in work hours and sleep filled with the training median. a blank practice quiz score always meant "no quizzes taken", so we added that as its own feature.
  • explanations: for each student we walk their path through the tree. at every split we compare the branch they took with the one they didn't: if theirs carries more risk, it becomes a reason with a matching action; otherwise it becomes a strength.
  • api: fastapi serves the model, the explanations, the evaluation metrics and the gemini study aids.
  • frontend: react, typescript, vite, tailwind css and recharts.

what it does

  • one real student, everywhere: the dashboard follows a student from the wolfhacks dataset. switch the student id and every page updates.
  • overview: a calendar where you add, tick off and remove deadlines and exams, with "due soon" counted from your list.
  • my progress: twelve numbers for the semester (midterm, attendance, on-time rate, study hours and more), each next to the average of the other students in the same course.
  • notes: upload a pdf, image or text file and get a summary, key points and a clickable practice quiz, generated by google gemini.
  • support check: a decision tree estimates whether a course needs "extra support", "some support" or is "on track", explains why in plain language, and gives a next step for each reason. change a number and check again to see what would move the result.
  • spookie's pond: a rubber duck in a witch hat for talking through what's stuck in a class. spookie knows what the dashboard shows for you, asks one question at a time, and ends with one small step.
  • resources: links to campus academic support, so a next step always has somewhere to go.
  • how the model works: recall, baseline comparison, the overfitting chart, the tree itself and its limits, on one page.

results

on the held-out rows the tree catches 89% of at-risk students (recall), with 49% precision and 72% accuracy. always predicting "not at risk" scores 73% accuracy but catches nobody.

we also compared against a one-line rule, "flag everyone with a midterm below 68". it catches 98% with 51% precision, so on this split it beats the tree. across 30 repeated splits the two roughly tie. we kept the tree because a midterm cutoff can only say "your midterm was low", while the tree points at attendance, deadlines and late-night studying, which a student can still change.

challenges

  • the tree did not clearly beat a simple rule. we decided to report that instead of hiding it.
  • recall versus false alarms. missing a struggling student costs more than a false alarm, so we scored recall twice as heavily as precision and lowered the cutoff to 0.20.
  • overfitting. recall drops from 97% on training rows to 89% on held-out rows.
  • gemini reliability. the api sometimes answered "high demand", so we added retries.
  • merging two people's work overnight. one push didn't compile after a formatter broke the markup, so we repaired it and connected the pages to the api in the morning.

responsible ai

  • supportive wording: results are shown as "on track", "some support" or "extra support", never as a prediction of failure, and always with strengths and next steps.
  • limitations: the data is synthetic, small and partly self-reported. the score is relative, not a calibrated probability, and the model has not been tested on real students.
  • privacy: uploaded notes are sent to google gemini to generate the summary and quiz. our api does not store or log them. no real student data is used.
  • ai disclosure: summaries and quizzes are generated by google gemini. parts of the code were written with claude code, as cited in the readme.

what we learned

a model that is easy to explain can be worth more to a student than one that scores slightly higher, and comparing against a strong baseline taught us more than the accuracy number did.

what's next

a weekly study plan built from your deadlines and the habits the model flags, and real course data in place of the example weekly plans.

Built With

Share this project:

Updates

Submission history