Inspiration
We noticed a frustrating reality in education: students spend hours reading textbooks and highlighting notes, but still fail exams. Research shows that passive reading has terrible retention rates. We were inspired by the Feynman Technique and Active Recall—proven cognitive science methods that force you to explain concepts in your own words. However, most students don't have a personal tutor to guide them through this process. We wanted to build an AI that doesn't just summarize text for you, but actually teaches you by challenging your understanding.
What it does
Synapse AI transforms passive reading into active learning in three steps: Visualize: Users upload any educational PDF. Our AI extracts the core concepts and maps their relationships, instantly generating an interactive, draggable Knowledge Graph. Interact: Users can click on any connection between concepts to enter Sage Mode 🌿. Learn: Sage is a Socratic AI tutor. Instead of giving the student the answer, it asks guiding questions, uses analogies, and gently corrects misconceptions to ensure true understanding. It also features an Adaptive Difficulty Slider (tuning the AI's vocabulary from "Explain Like I'm 5" to "Expert") and a Session History to revisit past graphs.
How we built it
We built a decoupled, full-stack architecture: Backend: Python and FastAPI. We use PyPDF2 to extract text from uploads, and the Google Gemini API to power both the Knowledge Graph extraction and the Sage Mode conversational tutor. Frontend: Next.js and ReactFlow for the interactive graph visualization, styled with Tailwind CSS for a polished, dark-mode-ready UI. Deployment: Both the backend and frontend are deployed live on Render.
Challenges we ran into
Prompt Engineering the Socratic Tutor: Getting an LLM to act like a teacher instead of a search engine was difficult. We had to heavily iterate on the system prompt to strictly forbid the AI from giving direct answers and force it to always end its replies with a guiding question. JSON Reliability: AI models sometimes wrap JSON in markdown blocks or add conversational filler. We had to write robust regex cleanup logic to ensure the frontend graph wouldn't crash on malformed responses. Full-Stack Integration: Managing CORS policies, environment variables, and asynchronous state between the React frontend and the Python backend during the deployment phase.
Accomplishments that we're proud of
We successfully built a true Socratic AI tutor. It is incredibly satisfying to watch the AI catch a student's misconception (e.g., "Sunlight is the food plants eat") and gently guide them to the right answer using a clever analogy. We are also proud of shipping a fully deployed, polished product with a custom UI, persistent local history, and adaptive learning in such a short timeframe.
What we learned
We learned how to effectively chain LLM prompts—using one prompt for strict structural data extraction (the graph) and another for open-ended conversational tutoring (Sage). We also learned the importance of UX in EdTech; building a smart AI is easy, but wrapping it in an intuitive, frictionless interface is what actually changes user behavior. On the technical side, we gained deep experience integrating ReactFlow with a FastAPI backend.
What's next for synapse-hackathon
Multi-Modal Input: Allowing users to upload photos of whiteboards or paste YouTube video URLs to generate graphs. Anki/Flashcard Export: Automatically turning the generated graph edges into spaced-repetition flashcards for long-term memory. Voice Mode: Letting students speak their answers to Sage for a more natural, hands-free tutoring experience. Cloud Sync: Adding user accounts so students can save their graphs and share study guides with classmates.
Log in or sign up for Devpost to join the conversation.