MemorAI_zer

About the Project

MemorAI_zer is more than just another flashcard app it’s the culmination of cognitive science research, AI engineering, and full-stack development. Here’s our journey…

🌱 Inspiration

The Forgetting Curve
Two weeks before my own exams, I’d cram all night only to forget 90% of it. I dove into Ebbinghaus’s forgetting curve and realized that without strategic review, our brains rapidly discard new knowledge.
Science of Learning
Reading Meyer (2024) on memory retention and Kosslyn (2017) on active recall and dual coding, I saw an opportunity to build something smarter than static “Anki‐style” decks.
AI’s Promise
Papers like Tabibian et al. (2019) showed AI-optimized spaced repetition outperforms fixed intervals. I wanted to harness Transformers, Deep Reinforcement Learning, and Bayesian Knowledge Tracing to create a truly adaptive learning companion.

🛠️ How I Built It

Backend (Flask + AI)
- Question Generation: Leveraged Hugging Face’s T5 (t5-small-e2e-qg) for end-to-end QG, wrapping it in a Flask API.
- Spaced-Repetition Scheduler:
  - SM-2 Fallback: Integrated supermemo2 for proven interval calculations.
  - Bayesian Knowledge Tracing: Used pyBKT to model mastery probabilities, dynamically updating after each recall.
Frontend (React + PWA)
Added User Accounts to sync progress across devices via a lightweight database (e.g., Firebase or PostgreSQL).
- Clean UI: Adopted a clean, responsive design with React, React-Bootstrap, and Framer Motion for smooth transitions.
- Service Worker & IndexedDB: Enabled offline functionality and local progress saving via a custom sw.js and idb.
- CORS & Proxy Setup: Configured the React dev server proxy and Flask-CORS to seamlessly bridge ports 3000↔5000.
Project Structure & Tooling
- Version Control: Organized into /backend (Python) and /frontend (React) folders, each with its own dependencies.
- Automated Debugging: Pinpointed errors—Hugging Face tokenizer conversions, pyBKT column mismatches, service worker 404s—using detailed logs and Google/StackOverflow research.

📚 What I Learned

Integrating Transformers Offline: Tackled fast vs. slow tokenizer issues by forcing SentencePiece with use_fast=False and handling tiktoken/blobfile dependencies.
Spaced Repetition Theory in Code: Merged classic SM-2, Bayesian models, and a DRL scaffold to compare and eventually blend multiple scheduling strategies.
PWA Nuances: Learned that service workers must live in the public/ folder, and how to correctly scope and unregister them during development.
Cross-Origin Workflows: Mastered CORS configuration in Flask and proxying in Create-React-App to avoid “Failed to fetch” errors.
Debugging Legacy Warnings: Updated deprecated Pandas .append() calls to pd.concat() and adapted to pyBKT’s expected column names (order_id, skill_name, etc.).

⚔️ Challenges Faced

Tokenizer Conversion Errors: Transformers’ slow→fast tokenization often broke on Windows; the workaround required deep dives into Hugging Face’s caching and conversion code.
pyBKT Data Formatting: Ensuring pandas DataFrames matched pyBKT’s strict column naming and non-empty schema took multiple iterations.
Service Worker 404s: Initially served sw.js from the wrong directory, leading to confusing “bad HTTP response code” logs.
CORS & Proxy: Balancing local development convenience (CRA proxy) against production readiness taught me the importance of environment-specific configuration.
Balancing Complexity: Integrating advanced AI models without over-engineering—keeping the code maintainable and hackathon-ready.

🎯 Next Steps

DRL Stub: Built a custom Gym environment and a Stable-Baselines DQN agent scaffold for future training.
Offline Imagery: Add a Stable Diffusion pipeline to auto-generate mnemonic images for each flashcard.
Train the DRL Scheduler on real recall data to compare against SM-2.
Enhance Dual Coding by automatically pairing generated images with flashcards.

Building MemorAI_zer was a deep dive into both the science of how we learn and the art of engineering robust, user-friendly systems. It challenged me to merge research papers with production code—and the result is a foundation for smarter, more personalized learning tools!