HASH# AI SCHOLAR

Home page for the application.
Generated summary from the sample 1 page PDF about LLMs.
Generated equations (none in current PDF) and explanations.
Generated method explanations.
Generated related papers and citations.
Generated conference insights from video transcripts. (1 min video on LLMs taken from youtube).

Inspiration

Reading research papers are often confusing and touch and time-taking. This is a common problem that students usually face and might want to save time in doing this.

What it does

This project doesn't just extract the summary of a research paper, but also extracts methods, identifies key findings, and suggests related works. This isn't a regular document summarizer, it also adds the human insights using information from the conferences where this paper has been presented/mentioned/talked about. This enriches the overall value the AI Scholar Assistant can add to any student's attempt to understand research papers.

How we built it

We built HASH# AI Scholar using a powerful combination of AI-driven natural language processing, document parsing, and interactive web technologies, all wrapped into a seamless Streamlit experience. Frontend & UI: Built with Streamlit, providing a responsive, and interactive interface for researchers to upload PDFs and transcripts, explore summaries, equations, and AI insights in real time. Text Extraction Layer: PyPDF2 and custom parsing logic in utils/pdf_parser.py for extracting clean, structured text from academic PDFs. Transcript parsing handled by utils/transcript_parser.py, enabling multi-format input. Summarization & NLP Core: Implemented with OpenAI GPT models (v1 API) and Transformers (Hugging Face) for context-aware summarization, section-wise breakdown, and conference insights. The summarizer intelligently condenses dense academic writing into concise, human-readable explanations. Mathematical & Method Analysis: Regex-based and transformer-assisted equation extraction from complex LaTeX patterns. The method explainer module uses LLM reasoning to interpret research methods and experimental design. Voice Narration: Integrated Elevenlabs to automatically narrate summaries for accessibility and auditory learners. Recommendation Engine: Related paper suggestions powered by semantic similarity embeddings using OpenAI text embeddings API and cosine similarity scoring. Deployment: Hosted on Streamlit Cloud, leveraging containerized Python environments with dependency management via requirements.txt. Fully serverless—users only need a browser to access AI Scholar.

Challenges we ran into

PDF Parsing Complexity: Research papers often have non-linear layouts (columns, equations, tables), making accurate text extraction and segmentation challenging. Token Management: Managing large document contexts with limited model input size required intelligent chunking and summarization strategies. We could only parese 10000 with our OpenAI API key, so, we had to use a 1 page PDF and 1 page transcript text file. Latency & Performance: Generating AI outputs (summaries, related papers, explanations) while keeping the interface responsive involved caching. Audio Integration: Ensuring smooth playback of dynamically generated mp3 across browsers required custom HTML embedding and base64 encoding.

Accomplishments that we're proud of

Attempted to build a fully functional AI research assistant that can read, summarize, explain, and even speak research papers. Designed a modular architecture (utils/ package) allowing easy extension and maintenance. Tried to develop a clean, professional Streamlit interface that feels like a scholar’s digital companion. Combined multiple AI capabilities - summarization, explanation, recommendation, and narration - into a single cohesive app.

What we learned

Latest trends and technologies in gen AI, including RAG based agentic AI, prompt engineering, cross-ref API, semantic vector search and generation, BERT, LLMs, MathPix, Evelenlabs API and streamlit.