dashboard

Inspiration

Students waste hours manually creating flashcards while generic AI chatbots hallucinate answers that contradict their actual class notes. We wanted to fix both — instantly.

What it does

SnapStudy converts any handwritten note, PDF, or text file into type-aware flashcards, an adaptive quiz, and a RAG-powered chat assistant — all grounded strictly in the student's own uploaded notes. Zero hallucinations. Zero manual effort.

How we built it

Frontend: React 18 + Vite + Tailwind CSS, deployed as a PWA on Vercel
Backend: Node.js + Express with 5 dedicated AI endpoints
AI: Google Gemini API — gemini-3.5-flash for unified single-pass note ingestion (OCR + flashcard generation in 1 API call), gemini-3.5-flash-lite for real-time RAG chat and remediation
Database: Supabase Postgres (JSONB regions + cards) + Supabase Storage bucket for note images, with Firebase fallback
Hallucination control: Strict JSON schema enforcement (responseMimeType: application/json, temperature: 0.2) + note context injection in every prompt

Challenges we ran into

Reducing Gemini API calls — solved by merging region detection + card generation into a single unified prompt (2–3× fewer calls)
Preventing LLM hallucinations — solved by injecting OCR'd note text as a bounded context block and physically cropping note regions with Sharp.js for grounded remediation
Multi-key rate-limit management — built an automatic API key pool failover system across 6+ keys

Accomplishments that we're proud of

Zero hallucinations by architecture — the RAG chat literally cannot answer outside the student's notes
Single-pass ingestion reducing API overhead by 50–67% per upload
Full PWA — installable, offline-capable, works on any device including $50 Android phones
Supabase integration with Row Level Security ensuring complete data isolation per user

What we learned

Prompt engineering at temperature: 0.2 with responseMimeType: application/json is the most reliable way to eliminate LLM drift
Client-side HTML5 Canvas compression (80–90% size reduction) is critical before any multimodal API call
RAG grounding is only as strong as the quality of the retrieved context — OCR region structure matters enormously

What's next for SnapStudy

Spaced repetition scheduling based on quiz session history already stored in Supabase
Teacher portal to share note decks with entire classrooms
Multi-language handwriting support and LaTeX formula rendering
Analytics dashboard showing which concepts students struggle with most