CosmosPapers: Universe of Research

An AI agent powered by Microsoft Copilot that helps researchers instantly discover, analyze, and summarize top tier conference papers helping in hours of tedious literature review.

Comment

Inspiration

Every year, over 10,000 papers are accepted across top-tier Computer Science conferences such as CVPR, NeurIPS, and ICLR. For professors, PhD students, and researchers, conducting a literature review means manually downloading hundreds of complex PDFs, reading dense abstracts, and manually formatting citations. The current workflow is overwhelming, fragmented, and massively time-consuming. We wanted to build a platform that didn't just aggregate these papers, but actually provided the intelligence needed to cut through the noise. We realized that by combining semantic search with an autonomous AI Agent, we could turn hours of tedious literature review into a conversation that takes seconds.

What it does

CosmosPapers is an AI-powered discovery platform with 50,000+ top CS papers. Features include:

Microsoft AI Agent: A slide-out Copilot that instantly summarizes papers and generates BibTeX citations. Semantic Search: Find papers by meaning, not just keywords, using pgvector. Top Papers Filter: Instantly find "Oral" and "Spotlight" award winners. Interactive Trends: Visualizations (stream graphs, bar races) showing topic evolution over time.

How we built it

AI Agent: Microsoft Azure OpenAI powered Copilot. Frontend: Next.js and Tailwind CSS with Recharts. Backend: FastAPI (Python) on an Oracle Cloud VM. Data: PostgreSQL with pgvector for instant semantic search. ML: BERTopic run on HPC clusters for macro-trend analysis.

Challenges we ran into

Managing 50,000 high-dimensional vectors for fast semantic search. Building a seamless AI Agent UI that doesn't force users to switch tabs. Writing resilient scrapers to standardize data from 11 differently formatted conference websites.

Accomplishments that we're proud of

We successfully integrated a specialized Microsoft Copilot Agent into a real-world workflow. It feels like a true research assistant, loading instantly and drastically cutting down the time needed to review complex literature.

What we learned

We learned how to optimize vector databases for speed, bridge complex ML backends (BERTopic) with a snappy Next.js frontend, and how to effectively deploy Microsoft's AI capabilities as the core value driver of an app.

What's next for CosmosPapers: Universe of Research

Ingesting full PDF texts (beyond abstracts) so the Microsoft Agent can answer hyper-specific questions about datasets and evaluation metrics. We also plan to roll out automated researcher profiles.

Built With

Share this project:

Updates