Inspiration
As graduate students, we've personally experienced the tough reality of academic research. Conducting a thorough literature review, which involves finding connections, tracking how methods develop across different papers, and synthesizing scattered information, is a monumental task that can take weeks or even months. Faced with over 5 million new papers published each year, researchers are compelled to go through the repetitive and inefficient cycle of manually reading each document, searching for insights often hidden deep within the text. This bottleneck not only wastes time but also slows down the overall pace of discovery.
What it does
Lumina transforms the painstaking process of reviewing academic literature into a dynamic, interactive conversation with your research. Instead of spending weeks manually sifting through hundreds of PDFs, a researcher simply uploads their papers and asks a natural language question, like, "Extract the sample size, methodology, and key outcomes from these studies." Our AI-powered agent then instantly creates a structured, queryable database tailored to your needs, removing the boring and error-prone task of manual data extraction. The real innovation, however, is the interactive layer: you can have conversations with your entire library. Ask complex questions such as, "Find all papers that critique this theory," and then follow up with, "For those papers, show me the original text where the authors state their main criticisms."
Lumina turns a static folder of documents into a dynamic, intelligent knowledge base, enabling researchers to synthesize information, recognize connections, and discover new insights with remarkable speed and accuracy.
How we built it
We developed a hybrid AI architecture by combining the efficiency of open-source models like Qwen3 (0.6 billion parameter) for fast parsing tasks with the advanced reasoning and synthesis abilities of powerful closed-source models like Gemini 2.5 Flash. For the front end, we created a user-friendly and accessible interface using Gradio, enabling rapid prototyping and providing a seamless, intuitive experience for researchers.
Challenges we ran into
Optimizing Local Model Deployment: We successfully configured and ran local, open-source models on resource-constrained hardware (Apple Silicon Macs), which required deep optimization of our inference stack. Integrating Disparate AI Frameworks: A key challenge was connecting local models with the OpenAI Agent SDK, which involved creating custom wrappers and modifications to ensure smooth communication and agent control. Overcoming SDK Limitations for Structured Data: We faced and developed a workaround for a known issue in the agent SDK that prevents simultaneous structured output (JSON) generation and tool calling. Our solution guarantees reliable data structuring without reducing agent functionality. Architecting for Scalability: Handling multiple, simultaneous document processing requests required building a robust asynchronous processing pipeline to manage connections and workflows efficiently without blocking the user interface.
Accomplishments that we're proud of
Scalable Multi-Document Analysis: Our system reliably parses and synthesizes insights from a large corpus of documents at once, addressing a common limitation of existing large context window models that often struggle with recall and accuracy when handling multiple files. Novel Agentic Framework for Data Structuring: We designed and implemented a sophisticated agentic AI framework that not only performs complex data extraction but also includes a validation loop, ensuring the accuracy and integrity of the generated structured data.
What we learned
Advanced Agentic AI Implementation: We gained extensive expertise in the OpenAI Agents SDK, mastering the creation of tool-using agents, decision-making logic, and the implementation of sequential, multi-step AI workflows. AI Self-Correction and Validation: A key learning was developing a robust validation loop for our models, which is essential for ensuring the reliability and accuracy of AI-generated structured data. Asynchronous Systems Architecture: We learned how to design and manage asynchronous request handling, a vital skill for building responsive and scalable applications that can process multiple documents and agent tasks in parallel. Rapid UI/UX Prototyping with Gradio: We quickly mastered Gradio for front-end development, learning how to swiftly build and deploy an intuitive user interface for a complex AI application.
What's next for Lumina
Persistent Knowledge Bases: Our immediate next step is to integrate a SQL database, allowing users to save, manage, and query their extracted data over time, turning one-off analysis into a persistent, evolving knowledge asset. Deep Insight Discovery with Knowledge Graphs: We plan to leverage Neo4j to build dynamic knowledge graphs from the extracted data. This will unlock a new level of insight, enabling users to visualize connections between papers, track the evolution of concepts, and discover hidden trends in their research field in a transparent and auditable way.
Log in or sign up for Devpost to join the conversation.