An AI Based Approach for Research Paper Summarization Using Deep Learning
TL;DR
The Research AI Pipeline is introduced, a five-layer automated system for analysing academic PDF documents from start to finish that achieves an F1-score of 0.89 for section boundary detection, 0.91 for figure caption matching and 88 percent accuracy in factual QA, all while operating locally without relying on cloud APIs.
Abstract
The growing number of scientific publications requires tools that can automatically convert raw PDF research papers into structured, searchable knowledge. This paper introduces the Research AI Pipeline, a five-layer automated system for analysing academic PDF documents from start to finish. Layer 1 ingests PDFs through a FastAPI gateway that streams data in real time. Layer 2 extracts text using a combination of pdfplumber, PyMuPDF and pytesseract to create a normalised JSON schema that includes sections, tables, figures and numerical data. Layer 3 conducts a thorough quality audit with thirty-two checks across six validation groups, producing a severity score on a one-hundred-point scale. Layer 4 employs local large language model reasoning using DeepSeek-V3 via Ollama, with prompts tailored to maintain numerical accuracy in six types of summarised sections and a final synthesis. Layer 5 organises the output and supports an interactive QA interface that streams responses based on the extracted content. Tested on fifty research papers across various fields, the system achieves an F1-score of 0.89 for section boundary detection, 0.91 for figure caption matching and 88 percent accuracy in factual QA, all while operating locally without relying on cloud APIs. The average processing time is 87 seconds per paper.