EvidenceDesk: Making AI-Powered Research More Evidence-Driven

Inspiration

Research today is easier to start than ever, but finding trustworthy answers is still difficult. AI can explain complex topics in seconds, yet a fluent, confident response does not automatically mean the information is accurate, the sources are relevant, or the evidence supports the conclusion.

For students and independent researchers, verifying information often means switching between search engines, research papers, websites, and AI tools. This creates a gap between getting an answer and understanding the evidence behind it.

EvidenceDesk was built around one central question: What if AI could help people investigate a question, rather than simply generate an answer?

Our goal was to make research more accessible, structured, and evidence-driven by bringing AI assistance and source-oriented research into one application.

What We Built

EvidenceDesk is an AI-powered research assistant built with Streamlit. It helps users explore research questions, discover relevant information, and investigate topics with an emphasis on sources and supporting evidence.

Instead of treating AI as an unquestionable authority, EvidenceDesk is designed to make AI-assisted research a starting point for further investigation. It aims to help users move from a question to a more informed understanding of the available information.

The project focuses on making the research process easier to navigate while keeping evidence and source discovery central to the experience.

How We Built It

We developed EvidenceDesk as a web application using Streamlit, allowing us to turn the research workflow into an accessible interactive interface.

The application combines AI-powered question answering with research-source retrieval. The research workflow uses external research data sources, including OpenAlex for scholarly literature discovery, alongside an AI model to help interpret questions and present useful responses.

The main development challenge was bringing these components together into a coherent workflow: understanding what a user wants to investigate, retrieving relevant information, and presenting the results in a way that supports further exploration.

Building the application also required attention to configuration, API integration, dependencies, and deployment so that the project could run beyond a local development environment.

Challenges We Faced

One of the biggest challenges was recognizing that generating a convincing answer and producing a useful research result are not the same thing.

Relevant sources do not automatically guarantee that every statement in an answer is correct. Search results can be incomplete, sources may address only part of a question, and AI-generated explanations can sound more certain than the available evidence warrants.

We also had to work through the practical challenges of connecting APIs, managing application configuration, debugging deployment issues, and making the application accessible through a working web interface.

These challenges shaped our approach: research tools should help users investigate information, not encourage them to accept generated responses without scrutiny.

What We Learned

Building EvidenceDesk strengthened our understanding of how AI applications combine language models, information retrieval, user interfaces, and external data sources.

More importantly, it taught us to think critically about the relationship between AI-generated answers and the evidence behind them. A useful research assistant should do more than summarize information; it should help users understand where information comes from and what still needs to be checked.

We also learned that turning an idea into a usable application involves much more than writing the core logic. Integration, testing, deployment, and communicating the project's purpose are all essential parts of building a functional product.

What's Next

EvidenceDesk represents a step toward a more transparent and accessible research experience. Future improvements could include stronger claim-level evidence mapping, clearer source attribution, better handling of conflicting findings, and more systematic evaluation of whether retrieved sources actually support generated claims.

Our long-term vision is to help students and researchers spend less time navigating fragmented information and more time understanding, questioning, and evaluating it.

EvidenceDesk is built on a simple principle: AI can accelerate research, but evidence should remain at the center of it.

Built With

• github
• machine-learning
• nvidia
• python
• research
• streamlit