Micro

Adaptive AI study platform that transforms unstructured course material into a personalized, continuously adapting study system.

Comment

Inspiration

Learning today is still built around static PDFs, linear textbooks, and one‑size‑fits‑all study plans. Students spend more time figuring out what to study than actually learning. We wanted to build something that feels like a personal tutor, one that understands your strengths, weaknesses, pace, and confusion points, and adapts instantly. Micro was inspired by three ideas:

  • Students shouldn’t need to manually organize or interpret course material.
  • Mastery should be measurable, probabilistic, and continuously updated.
  • Study plans should adapt in real time, just like a great human tutor would. We asked: What if your course PDFs, slides, and notes could transform themselves into a personalized, intelligent study system?

What it does

Adaptive AI study platform that transforms unstructured course material into a personalized, continuously adapting study system. Upload PDFs, lecture slides, and notes. Micro ingests them, builds a structured topic tree, models what we know using a Bayesian belief system, and drives every subsequent decision, what to study, in what order, for how long, and how.

How we built it

Micro is a full-stack system combining AI, statistical modeling, embeddings, and a modern frontend. Content Ingestion + Topic Tree

  • Extracted text from PDFs, slides, and notes.
  • Used pgvector embeddings to cluster concepts into a hierarchical topic tree.
  • Linked extracted topics to mastery models and study actions. Bayesian Mastery Modeling (Statistical Learning Model)
  • Built a Bayesian Knowledge Tracking engine using a Beta–Bernoulli model.
  • Designed confidence‑ and speed‑weighted updates: $$\alpha \leftarrow \alpha +w_t,\quad \beta \leftarrow \beta +w_t$$
  • Computed mastery as the posterior mean: $$\hat {\theta }=\frac{\alpha }{\alpha +\beta }$$
  • Modeled uncertainty using posterior variance and credible intervals.
  • Added exact mastery probability outputs per topic and per student.
  • Integrated mastery into the study plan engine, powering recommendations and progress indicators.

Frontend

  • React 19 + Vite 7 + Tailwind CSS 4.
  • Dashboard with mastery‑driven progress bars and suggested study plan.
  • Course view with interactive mindmap, content browser, and mastery indicators.
  • AI Panel with Ask/Act/Advise modes.

Backend + Database

  • Supabase PostgreSQL with 28‑table schema.
  • pgvector embeddings for topic clustering.
  • RLS policies for secure multi-user access.
  • Realtime updates for study sessions and mastery changes.

Challenges we ran into

  • Converting messy, unstructured academic content into a clean, hierarchical topic tree.
  • Designing a mastery model that is both interpretable and mathematically sound.
  • Balancing Bayesian stability with responsiveness to new evidence.
  • Ensuring the UI updates instantly as mastery changes.
  • Managing embeddings, file storage, and real-time updates in a single Supabase schema.
  • Creating a unified dev server that runs both the API and frontend seamlessly.

Accomplishments that we're proud of

  • Built a fully functional adaptive learning system in under 23 hours.
  • Designed a statistical learning model that outputs exact mastery probabilities.
  • Created a topic tree generator that works on arbitrary PDFs and notes.
  • Integrated embeddings, Bayesian inference, and React UI into one cohesive system.
  • Achieved a smooth, VS Code–inspired UI with real-time progress updates.
  • Delivered a system that feels like a personal tutor, not just another flashcard app.

What we learned

  • How to combine statistical modeling with modern LLMs and embeddings.
  • How to design mastery systems that are interpretable, stable, and mathematically grounded.
  • How to build a full-stack AI product with React, Supabase, and custom Python models.
  • How to structure academic content into machine‑understandable topic graphs.
  • How to design adaptive study plans that respond instantly to user behavior.

What's next for Micro

  • Integrate the Bayesian mastery model directly into the production engine (it currently runs only in the research notebook).
  • Connect Bayesian mastery to the topic tree, so every concept has a live belief state.
  • Use mastery probabilities to drive Next Up, difficulty selection, pacing, and review scheduling.
  • Add Deep Knowledge Tracing (DKT) as a second layer for sequence‑aware learning patterns.
  • Expand ingestion to diagrams, handwritten notes, and multimodal slides.
  • Build instructor dashboards and cohort‑level mastery analytics.
  • Ship a mobile app with offline study mode.

Built With

Share this project:

Updates