StudyPack Tutor - Offline Classroom Copilot

Inspiration

Teachers and students often work with limited or no internet, but still need high-quality explanations, lessons, and quick quizzes that respect privacy. I wanted a “local-first” tutor that answers only from the materials I provide, so it stays on-topic, works offline, and never leaks data.

What it does

StudyPack Tutor is an offline classroom copilot.

• Loads your study packs (.pdf, .txt, .md) by folder.
• Answers questions using only those materials (RAG).
• Generates lesson plans with objectives, steps, and exit tickets.
• Creates quizzes and auto-grades student answers with concise feedback.
• Optional bilingual output + adjustable reading level for age-appropriate responses.

How we built it

• Backend (Flask): endpoints for ask/stream, lesson, quiz, and grading.
• Local LLM via Ollama: chat completions against gpt-oss:20b.
• RAG indexer: TF-IDF + cosine similarity over chunked text from .pdf/.txt/.md (PyPDF2 + scikit-learn).
• Frontend (Streamlit): tabs for Tutor Chat, Lesson Generator, Quiz + Auto-Grader, with inline instructions, preset prompts, and a pack picker.
• Safety: simple keyword guardrails for inappropriate content.

Challenges we ran into

• Formatting: streamed text arrived as a blob; needed a Markdown normalizer to force bullets/headings.
• Windows streaming: Flask reloader/threads could drop connections, fixed by debug=False and use_reloader=False.
• PDF text quality: some PDFs extract poorly; added support for .txt and .md to make RAG robust.
• Clear UX: users were unsure why “full moon date” can’t be answered offline, added inline help + presets and visible pack selection.

Accomplishments that we're proud of

• Fully offline assistant with clean, simple deployment.
• Consistent, source-grounded answers that don’t hallucinate beyond user files.
• Turn-key lesson and quiz generation that teachers can actually use.
• Bilingual + grade-level controls that make outputs teacher-ready.

What we learned

• Small, obvious UX nudges (help cards, presets, and token sliders) dramatically improve success rates.
• A lightweight TF-IDF retriever is often enough for K-12 content if chunking and prompts are done carefully.
• “Local-first” constraints force better engineering: predictable latencies, smaller outputs, and resilient streaming.

What's next for StudyPack Tutor - Offline Classroom Copilot

• Drag-and-drop uploader inside the UI (auto-reindex on change).
• Per-answer citations UI (click to reveal exact snippet).
• Pack manager (rename/delete packs, show file counts).
• On-device embeddings to replace TF-IDF for better recall.
• Export lessons/quizzes to DOCX/PDF and Google Classroom-ready CSV.

Team

Solo: Sweety Seelam

Built With

• flask(api)
• gpt-oss:20b-model
• http-requests
• ollama
• pypdf2
• python
• scikit-learn
• streamlit(ui)