Inspiration

Passive reading creates the illusion of competence—students re-read 50-page PDFs and still blank on exams. While most AI tools stop at summaries, passive recognition rarely builds long-term retention. I built Recalyn to convert static course PDFs into active-recall testing instruments with live retention telemetry.

What it does

Recalyn (recalyn.apk) transforms multi-page PDFs into schema-validated flashcards, timed quizzes, and real-time mastery tracking. Learners flip cards in a tactile Risograph deck, filter sessions by #TAG, review weak cards, and take Fisher–Yates randomized 30-second quizzes that prevent position memorization. Progress updates a live 0–100 composite course grade: $$G = 0.70 \cdot Q_{\text{weighted}} + 0.20 \cdot C_{\text{retention}} + 0.10 \cdot S_{\text{streak}}$$ where $Q_{\text{weighted}}$ is difficulty-weighted quiz accuracy, $C_{\text{retention}}$ is flashcard recall confidence, and $S_{\text{streak}} = \min(\text{streak}_{\text{days}} \times 20,\ 100)$ rewards daily consistency.

How we built it

The mobile client uses React Native (Expo SDK 52), TypeScript, and react-native-reanimated with a custom two-ink Risograph UI, packaged via EAS Build. The backend runs Python 3.12, FastAPI, and SQLAlchemy with PyMuPDF ($45{,}000$-character page-bounded extraction) and Gemini AI enforced by Pydantic v2 JSON schemas. Deployed via Docker on Render with PostgreSQL and GitHub Actions CI/keep-alive automation.

Challenges we ran into

Android’s document picker returns content:// URIs rejected by native fetch(), solved via client-side XMLHttpRequest blob-to-base64 encoding. Migrating to PostgreSQL triggered a psycopg v3 dialect mismatch, resolved by normalizing connection URLs to postgresql+psycopg://. I also built an AsyncStorage runtime API switcher with /health probes to handle free-tier cold starts.

Accomplishments that we're proud of

Shipping an end-to-end Android APK backed by containerized FastAPI and PostgreSQL on $0/month infrastructure, achieving zero malformed AI outputs via Pydantic schemas, and delivering a fast, tactile study experience.

What we learned

Strict schema validation makes LLM outputs deterministic enough to drive grading math, and mobile networking requires defensive endpoint switching and health diagnostics.

What's next for Recalyn

Adding SM-2 spaced repetition scheduling, offline AsyncStorage sync queues, and lecture slide or transcript ingestion.

Built With

Share this project:

Updates

Submission history