Skip to content

About

An intelligent, personalized learning system that evaluates student answers using NLP, identifies weak concepts, estimates concept mastery, dynamically adjusts question difficulty, recommends the next question, and tracks learning progress.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

Study Grove — Adaptive AI Tutor

A retro-inspired, forest-green and sage-green adaptive tutor for the 18-question machine-learning question bank in Adaptive_AI_Tutor.ipynb.

The visual direction is a clean vintage study journal: warm paper, ruled notes, botanical green accents, serif editorial headings, small mono labels, and restrained borders rather than a glossy dashboard aesthetic.

Features

  • Adaptive questions across Overfitting, Classification, Regression, Cross Validation, Underfitting, and KNN.
  • Easy / Medium / Hard selection using the same mastery thresholds as the notebook.
  • Browser-based semantic answer scoring with Xenova/all-MiniLM-L6-v2 through Hugging Face Transformers.js. The model downloads the first time it is used and is cached by the browser where supported.
  • Clearly labelled keyword-overlap fallback if model loading/inference is unavailable.
  • Same answer labels as the notebook: score >= 70 = Correct, >= 45 = Partially Correct, otherwise Incorrect.
  • Same mastery adjustments: +10 / +3 / -10, clipped to 0–100.
  • 10-question sessions, concept mastery dashboard, weak-concept priorities, and recent attempt history.
  • Local browser persistence with JSON and CSV export, no account required.
  • Static-first Vercel deployment: no Python Function is needed for the website, which avoids serverless startup and static-file routing issues.

Important scoring note

This follows the notebook's approach: cosine similarity between the student's answer and the reference answer, multiplied by 100. That similarity is a heuristic, not a calibrated probability or a definitive judgement of learning. The fallback score is content-word F1 and is explicitly labelled. If a model load fails, ask learners to compare their answer with the reference answer instead of treating the score as authoritative.

Run locally

You can use a simple local HTTP server (a secure/served origin is required for modules and browser model loading):

python -m http.server 8000 --directory public

Then visit http://localhost:8000.

Deploy to Vercel

  1. Download and extract the project ZIP.
  2. Create an empty GitHub repository, for example adaptive-ai-tutor.
  3. Push the contents of this folder to the repository root. The root should contain vercel.json, public/, notebooks/, and this README.
  4. Import the repository at Vercel.
  5. Keep Root Directory pointed at the folder containing vercel.json.
  6. This is a static site: leave custom Build Command and Install Command blank. vercel.json sets outputDirectory to public.
  7. Deploy. Vercel should serve public/index.html as /, with /style.css, /app.js, /favicon.svg, and /data/question_bank.json at their root URL paths.

There is deliberately no app.py or serverless API required for the primary deployment. The complete tutor runs in the browser, so Vercel won't need to start a Python function. The Sentence Transformers model is downloaded from Hugging Face through a CDN, which means the semantic engine requires an internet connection the first time it is used.

Data storage and privacy

Progress is stored in localStorage in the current browser on the current device. It is not a cloud student account or a multi-device database. Clearing browser data, switching browsers/devices, or using private mode may reset it. Export JSON/CSV to keep a backup. Student-written text and progress are not sent to a project backend; model inference runs in the browser after the model download.

Colab notebook

  • notebooks/Adaptive_AI_Tutor.ipynb is a copy of the uploaded notebook.
  • notebooks/Adaptive_AI_Tutor_web_deploy.ipynb is a copy with an additional export cell that writes the question bank and scoring configuration to JSON and packages them for download.

The added export cell is self-contained. Open the deployment notebook in Colab and run only the final export cell if you want to regenerate the JSON data; you do not need to run the interactive input() tutoring session. If the original questions list is already defined in the notebook runtime, it uses that list; otherwise it uses an embedded copy of the original question bank.

Project structure

adaptive-ai-tutor-retro/
├── vercel.json
├── README.md
├── public/
│   ├── index.html
│   ├── style.css
│   ├── app.js
│   ├── favicon.svg
│   └── data/
│       ├── question_bank.json
│       └── model_config.json
└── notebooks/
    ├── Adaptive_AI_Tutor.ipynb
    └── Adaptive_AI_Tutor_web_deploy.ipynb

Credits / model

The tutor's question bank and mastery rules are taken from the uploaded Adaptive_AI_Tutor.ipynb. The browser uses the ONNX-compatible Xenova/all-MiniLM-L6-v2 model with Transformers.js to mirror the notebook's Sentence Transformers approach. See the Transformers.js pipelines documentation for model loading and feature-extraction behavior.

About

An intelligent, personalized learning system that evaluates student answers using NLP, identifies weak concepts, estimates concept mastery, dynamically adjusts question difficulty, recommends the next question, and tracks learning progress.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages