🧠 About ScholarSync
💡 Inspiration
The inspiration for ScholarSync came from a common struggle among the student community at Hack Days Pune: Research Fatigue.
When starting a project or a thesis, researchers are often buried under a mountain of PDFs ($P_1, P_2, \dots, P_n$). These papers often contain contradictory data or varying methodologies. Traditional AI tools use RAG (Retrieval-Augmented Generation), which slices these documents into small chunks. While efficient, this causes the AI to lose the "Global Context" $\mathcal{G}$, making it impossible to see how Paper A's methodology directly refutes Paper B's conclusions. We wanted to build a "War Room" that treats your entire library as a single, coherent brain.
🔍 What it does
ScholarSync is an agentic research tool that ingests multiple PDFs and utilizes the Gemini 1.5 Flash 1-million-token context window to perform a "Full-Spectrum Synthesis." It identifies:
Consensus: Common findings across all authors.
Conflict Map: Specific areas where $\text{Author}_A \neq \text{Author}_B$ regarding data or estimates.
Gaps: Areas in the field that remain unexplored by the current library.
🛠️ How we built it
The project is built on a modern Software-Only stack:
Backend: A Python-based Flask server acting as an orchestrator.
AI Engine: The new google-genai SDK (v1.61.0) was used to interface with Gemini 1.5 Flash. We utilized Strict JSON Mode to ensure the model's reasoning was always machine-readable for our UI.
Processing: PyPDF2 handles server-side extraction, transforming binary document trees into high-density text strings.
Frontend: A responsive, dashboard-style UI using Tailwind CSS and Vanilla JavaScript for asynchronous API communication.
We can represent the synthesis process mathematically as:
where $S$ is the final structured synthesis and $P_i$ represents each uploaded research paper.
🚧 Challenges we faced
The SDK Migration: One of our biggest hurdles was the sudden deprecation of the google.generativeai package. We had to pivot mid-hackathon to the new google.genai client, which required a complete rewrite of our API calls and configuration logic.
Context Density: Handling massive text strings from multiple 50-page PDFs can cause latency. We optimized our text extraction to strip non-essential whitespace and metadata to maximize the token efficiency of the prompt.
Structured Reasoning: Forcing an LLM to identify contradictions is harder than asking for a summary. We had to iterate heavily on our System Instructions to ensure the AI didn't just summarize, but actively "hunted" for conflicts.
🎓 What we learned
We learned that the true power of the next generation of LLMs isn't just "talking," but Reasoning over Scale. Developing ScholarSync taught us how to manage large-scale context windows and the importance of structured outputs (JSON) in building reliable agentic tools. Most importantly, we learned that in a hackathon, the ability to adapt to new SDKs and API changes is just as important as the initial idea.
Log in or sign up for Devpost to join the conversation.