EdSynapse

• 

  logo

💡 Inspiration

I kept watching my classmates use ChatGPT as a copy-paste shortcut to pass assignments without actually learning anything. I wanted to build a tutor that does the opposite: one that adapts to your learning style, grounds itself in your syllabus, and asks Socratic questions to make you think. I fell in love with this concept because it connected directly to my master's studies — so much so that I registered it as a real company: EdSynapse Educational Services Inc., and this hackathon became the launchpad.

⚙️ What it does

EdSynapse runs a Diagnose ➔ Teach + Check ➔ Verify loop. It features Course Mode (teachers upload syllabus materials and track student progress) and Self-Study Mode (students upload their own notes). The AI maps your strengths on a visual Knowledge Map, teaches you using only your uploaded files, and quizzes you to verify your understanding.

🛠️ How we built it

EdSynapse is a single full-stack Next.js 15 (App Router) app — UI scaffolded with Vercel v0 and Claude Code, deployed on Vercel. It's backed by Amazon Aurora PostgreSQL with the pgvector extension, which handles both the highly relational domain (users, courses, topics, knowledge maps, attempts) and vector retrieval for source material in one database. Uploaded material is chunked, embedded, and stored so every answer is RAG-grounded. The OpenAI API powers streaming tutor chat, quiz and assessment generation, grading, and embeddings.

🚧 Challenges faced

Passwordless Latency: Generating AWS IAM database tokens on the fly caused connection lag, so we had to build a custom token-caching pool. Database Wakeups: Aurora Serverless v2 scales down when idle, so we wrote a custom connection retry loop with backoff to handle wake-up timeouts. RAG Precision: Cosine search alone sometimes pulled irrelevant text. We built a two-stage RAG pipeline that pulls candidates with pgvector and uses a lightweight LLM to re-rank the most relevant passages.

🧠 What we learned

We learned that good teaching is about shifting the cognitive work back to the student, not giving them easy answers. On the tech side, we realized a single SQL database with pgvector is incredibly powerful and completely eliminates the need for a separate vector database.

Built With

• amazon-web-services
• aurora
• next.js
• openai
• pgvector
• postgresql
• react
• resend
• tailwind
• typescript
• v0
• vercel