Yumi

Yumi is a local-first campus study assistant built on a standard PyTorch stack. It is designed for offline deployment and local data residency.

Current Features

• Course and material management (local text ingestion + chunking)
• Material file upload (txt/md/pdf/image) with OCR for image notes
• Classroom audio intelligent assistant:
  • offline lecture transcription + summary
  • custom terminology normalization
  • simple speaker diarization (Speaker A/B)
  • Anki flashcard CSV export
  • batch processing with progress visualization
• Note summarization (summary + keywords)
• Local QA with source excerpts
• Final-week schedule generation
• Fixed-event conflict avoidance (classes, meetings, commute)
• Plan analytics (hours by day/course + load stability)
• ICS export for calendar apps

Tech Stack

• PyTorch + lightweight local NLP adapter (replaceable with larger local models)
• FastAPI for backend service
• Streamlit for local UI
• SQLite for local storage

Quick Start

1. Install dependencies

pip install -r requirements.txt

2. Initialize DB

python scripts/init_db.py

3. Run API

python scripts/run_api.py

4. Run UI (new terminal)

python scripts/run_ui.py

Default URLs:

• API: http://127.0.0.1:8000
• UI: http://127.0.0.1:8501

Mobile Use (Phone)

1. Ensure phone and computer are on the same Wi-Fi.
2. Start UI:

python scripts/run_ui.py

3. The script prints LAN address like http://192.168.x.x:8501.
4. Open that address on your phone browser.

Optional environment variables:

• YUMI_UI_HOST (default 0.0.0.0)
• YUMI_UI_PORT (default 8501)
• YUMI_API_HOST (default 127.0.0.1)
• YUMI_API_PORT (default 8000)
• YUMI_DEV_RELOAD (1 to enable API hot reload)

Native App (Android)

Android app shell is included at: mobile/yumi-android

It provides:

• installable APK
• server URL input and persistence
• full WebView access to Yumi
• file picker support for upload workflows

Build instructions: see mobile/yumi-android/README.md

Scheduling Logic

Course priority score:

priority = 0.35*urgency + 0.30*(1-mastery) + 0.20*difficulty + 0.15*credit_weight

Planner rules:

• Deep block: default 90 minutes
• Review block: default 30 minutes
• Buffer time: default 20%
• Mandatory reviews at D-7, D-3, D-1
• Fixed weekly events remove occupied intervals before planning

Core API Endpoints

• POST /courses
• GET /courses
• POST /courses/{course_id}/materials
• POST /courses/{course_id}/materials/upload
• POST /audio/process-upload
• GET /audio/transcripts?course_id=&limit=
• GET /audio/{transcript_id}/anki.csv
• POST /glossary/terms
• GET /glossary/terms
• POST /notes/summarize
• POST /qa/ask
• POST /planner/exams
• GET /planner/exams
• PUT /planner/availability
• GET /planner/availability
• POST /planner/fixed-events
• PUT /planner/fixed-events
• GET /planner/fixed-events
• POST /planner/final-week-plan
• GET /planner/events?start_date=YYYY-MM-DD&end_date=YYYY-MM-DD
• GET /planner/analysis?start_date=YYYY-MM-DD&end_date=YYYY-MM-DD
• GET /planner/export/ics?start_date=YYYY-MM-DD&end_date=YYYY-MM-DD&include_fixed=true

Database Tables

• courses
• notes
• document_chunks
• exams
• availability_slots
• fixed_events
• study_events

Next Extensions

• Higher-accuracy domain ASR model packs
• FAISS embedding retrieval
• Personal knowledge graph per course

Audio Assistant Positioning

• Problem solved:
  • students/teachers spend too much time turning lecture recordings into structured notes
  • manual transcription is slow and key points are hard to extract
• Core promise:
  • data stays on device from audio input to note output
  • no network requests during inference (except first-time model download if not cached)
• Typical scenarios:
  • lecture recording review
  • seminar archive
  • group discussion minutes
  • language listening recap
• Target users:
  • undergraduates (heavy coursework in engineering/medical majors)
  • postgraduates
  • teachers
  • self-learners

OCR Runtime Note

Image OCR uses pytesseract, which requires local Tesseract OCR binary installation. On Windows, install Tesseract and ensure tesseract.exe is in PATH.