Secure AI Applications for Summarizing Scientific Studies and Preserving Privacy

Abstract

The exponential growth of scientific literature presents a critical challenge for researchers: existing summarization tools force an unacceptable trade-off between factual accuracy and data security. Extractive methods preserve factual integrity but produce incoherent, disjointed summaries lacking academic narrative quality. Cloud-based abstractive methods using Large Language Models (LLMs) generate fluent text but require uploading potentially sensitive research data to external servers, raising serious concerns about privacy, cost, and dependency on third-party services. This paper introduces Abstractive Muse, a novel, self-contained, and privacy-preserving framework designed to eliminate this trade-off. The system is delivered as a standalone desktop application that integrates a classic graph-based algorithm, TextRank, for initial extractive summarization with a locally-executed, open-source LLM (Mistral-7B, ~2.6 GB in GGUF format) for abstractive synthesis, tested on a curated dataset of 10 peer-reviewed scientific PDF documents spanning domains including medicine, computer science, and engineering. We detail the complete development lifecycle, including the architectural design, the implementation of a user-centric Graphical User Interface (GUI), hallucination mitigation strategies, and the engineering decisions that ensure robustness and independence from external APIs. The final application empowers users to select a local PDF document, define summarization parameters, and generate not only a salient extractive summary but also a high-quality, AI-powered academic narrative, all without their data ever leaving their machine. This work presents a replicable blueprint for building practical, secure, and accessible NLP tools and argues for the value of local-first AI in academic research.