Inspiration
I decided to write a research paper myself and realized the critical bottleneck of information overload in academic research. Currently, researchers must manually search, read, and cross-reference dozens of dense technical papers to understand the state-of-the-art on a specific topic, making the literature review process highly inefficient and prone to human error.
What it does
Instead of a human spending weeks finding, reading, and summarizing complex scientific papers, your application does it automatically in minutes.
Here is exactly what it does, step-by-step:
- You ask a question: You type a complex scientific topic into the search bar (for example, "Find papers on deep learning for X-ray diagnosis").
- The Agents go to work: Instead of just one AI trying to do everything, your system uses a team of specialized AI agents:
- The Searcher Agent scours academic databases to find the actual research papers.
- The Analyst Agent reads those papers and pulls out the important math, data, and methodologies.
- The Synthesizer Agent takes all those separate notes and writes a clean, coherent summary.
- You get a final report: The frontend of your application displays a fully written, formatted literature review comparing the papers, saving the user hundreds of hours of manual reading.
Basically, it turns a massive pile of dense research papers into a single, easy-to-read summary document.
How we built it
The system was built using a decoupled client-server architecture to separate the user interface from the heavy processing logic. -The Frontend: We developed the client side using React.js and styled it with Tailwind CSS to ensure a responsive, clean interface tailored for researchers. It handles complex user inputs and provides real-time visual feedback on agent status. -The Backend: The core engine is built in Python. It acts as the orchestrator for the multi-agent framework. -The AI Layer: We engineered specific prompts and workflows to create specialized Large Language Model (LLM) agents—a Searcher, an Analyst, and a Synthesizer. These agents communicate with each other on the backend, processing the data sequentially before sending the finalized literature review back to the React frontend via REST APIs.
Challenges we ran into
1.Agent Orchestration and Context Limits: One of the hardest parts was ensuring that the LLM agents passed information to one another cleanly without hallucinating or losing highly technical data (like mathematical formulas or specific methodologies) during the handoff from the Searcher to the Analyst. 2.Asynchronous Processing: Multi-agent workflows take time to execute. Managing the state in React and ensuring the frontend did not time out while waiting for the backend Python engine to finish reading and synthesizing multiple papers was a significant technical hurdle. 3.Formatting Complex Output: Rendering dense STEM literature, including citations and academic structures, dynamically in the browser required careful parsing of the LLM output.
Accomplishments that we're proud of
True Agent Collaboration: We successfully built a system where multiple AI models do not just generate text independently, but actually collaborate, pass context, and refine each other's work to produce a single, cohesive academic document.
Bridging the Gap: We are proud of translating a highly complex backend architecture into a clean, intuitive web application that any researcher can use without needing to understand the underlying AI engineering.
Real-World Utility: Developing a final-year major project that directly solves a massive, real-world bottleneck in the academic and scientific community.
What we learned
Advanced LLM Architecture: We gained deep, practical insights into multi-agent orchestration, moving beyond simple chatbot implementations into structured, agentic workflows.
Full-Stack Integration: We learned how to effectively bridge a heavy Python processing backend with a dynamic React frontend, specifically handling long-running asynchronous tasks and state management.
Prompt Engineering for Logic: We learned that prompting an LLM to extract methodologies or critique a paper requires a very different approach than prompting an LLM to simply summarize text.
What's next for Multi-Agent System for STEM Literature Synthesis
WebSocket Integration: Transitioning from REST APIs to WebSocket connections to allow real-time, token-by-token streaming of the synthesized text directly to the frontend as the agents write it.
Citation Graph Visualizer: Building an interactive UI component that visually maps out the relationships and overlapping citations between the synthesized papers.
Authentication and Workspaces: Adding user accounts so researchers can save their literature reviews, revisit past syntheses, and build an ongoing library of AI-generated STEM insights.
Built With
- batchfile
- css
- javascript
- powershell
- python
- react
- typescript
