LecAssist / Aura

An academic-assistance thesis combining lecture understanding with agentic workflow experiments.

LecAssist is my final-year thesis at Sukkur IBA University (2026–2027). This workspace includes Aura, an academic recording assistant built with React, TypeScript, Electron, and FastAPI, plus academic portal-automation work.

Architecture

React + TypeScript / Electron
            ↓
         FastAPI
            ↓
Audio → faster-whisper → transcript → AI summary
                                      ↓
                               SQLite lecture library

The backend includes provider configuration for Ollama, OpenAI, and Claude. Lecture-library questions use stored summaries as model context. The current public library-question path does not use embedding-based vector retrieval.

Repository map

Path	Purpose
Aura/	Desktop client, Electron entry point, and Python backend
portal-automation/	Academic portal-automation work
DEMO/	Demo materials
transcripts/	Transcript materials

Development

git clone https://github.com/asadsehto/fyp.git
cd fyp/Aura
npm ci
python -m venv .venv

Activate the Python environment for your operating system, then install the backend dependencies:

python -m pip install -r backend/requirements.txt

Package scripts:

npm run dev           # Vite frontend
npm run electron:dev  # Vite + Electron
npm run build         # TypeScript check + frontend build
npm run lint          # Oxlint

The Python backend entry point is backend/server.py and listens on port 8000. When running it independently:

python backend/server.py

Check whether the Electron launcher has already started the backend before starting another instance. Audio libraries, model downloads, provider configuration, and external agent tooling are additional runtime requirements.

Research focus

• Grounded student question answering under constrained compute.
• Evaluation of multi-step agent trajectories when multiple action sequences may be valid.
• Explicit component failures rather than silently confident output.

Current limitations

This is an active thesis prototype. If Whisper cannot load its model, the backend can fall back to mock transcription; that output is not a real transcript. Automation and message delivery depend on configured external tooling. Local configuration is persisted to disk and should be reviewed before storing real account credentials.

Author

Asad Saleem · LinkedIn