DejaView: An AI-Powered Academic Project Integrity and Evaluation Portal Using Vector Embeddings and Large Language Models

Abstract: The proliferation of student project submissions in higher education institutions has introduced acute challenges in maintaining academic integrity and evaluation consistency. Manual review processes are labour-intensive, subjective, and fail to scale with the growing volume of submissions. This paper presents DejaView, a full-stack, AI-powered academic project management portal developed for the Department of Computer Science and Engineering (AI & ML) at Dayananda Sagar University. DejaView automates two critical pain-points: (1) detection of plagiarised or over-similar submissions using 384- dimensional dense vector cosine similarity powered by the FastEmbed BAAI/bge-small-en-v1.5 model, and (2) structured generation of executive summaries, technology stack labels, and personalised viva voce questions using the Groq LLM (llama- 3.3-70b-versatile). The system is built on a modular three-layer Python architecture comprising a Streamlit frontend, an AI sentinel engine, and a thread-safe SQLite persistence layer. Institutional access is enforced through strict email domain gating and bcrypt password hashing, ensuring that only verified university stakeholders interact with the platform. Experimental results demonstrate that the system reliably flags submissions with greater than 75% cosine similarity, generates domain-relevant AI insights with high contextual precision, and exports structured grade sheets for archival in a single click.

Keywords: Academic Integrity, Plagiarism Detection, Vector Embeddings, Large Language Models, Federated Evaluation, Streamlit, Cosine Similarity, Viva Question Generation, FastEmbed, Groq API.