Spark your Learning

AI can get to know you better than any teacher could. Let AI enable you to Learn and Grow

Comment

Inspiration

Mastering music theory—especially syllabus-heavy standards like ABRSM—is often intimidating, expensive, and rigid. Students struggle with abstract concepts (intervals, transpositions, time signatures), while traditional workbooks offer zero real-time feedback or audio reinforcement. Private 1-on-1 tutoring is inaccessible to many, and static online quizzes fail to adapt when a student is stuck on a foundational prerequisite.

I built SparcLearn to bridge this gap: an intelligent, adaptive music theory platform that combines strict musicology rules, dynamic notation rendering, and AI tutoring to deliver a truly personalized learning path for every student.

What it does

SparcLearn provides comprehensive, structured learning across ABRSM Grades 1–5:

  • Comprehensive Syllabi (Grades 1–4): Structured 16-week curricula with verified video lessons, interactive notation, and syllabus-scoped practice.
  • Adaptive Grade 5 Engine: A 24-question diagnostic placement quiz maps the student's mastery across 35 learning objectives and builds a personalized 6-week study schedule.
  • Computed Music Theory Engine & Unlimited Practice: 100% computed music theory (keys, intervals, chords, transposition) powering thousands of unique, dynamically generated exam-style questions with zero defects.
  • Interactive Engraved Notation & Web Audio: Interactive sheet music rendered via VexFlow with synchronized browser-based audio playback.
  • Grounded AI Tutor: A safe AI study companion that explains concepts in alternative ways, diagnoses why a student got an answer wrong, and answers deep-dive questions grounded in lesson context.
  • Mastery Tracking & Spaced Repetition: Real-time exponential mastery scoring with forgetting-curve decay to resurface decayed concepts before exam day.
  • Parent Dashboard & Pro Subscriptions: Full transparency for parents on student mastery and seamless Stripe checkout for Pro features.

How I built it

  • Frontend: Vanilla JavaScript (ES6+), HTML5, and CSS3 without heavy framework overhead for lightning-fast sub-second loading; VexFlow for dynamic music engraving; Web Audio API for synthesized multi-voice playback.
  • Backend & Persistence: Lightweight Python stdlib HTTP service (server.py) backed by SQLite (WAL mode) with atomic JSON/state transactions and signed stateless sessions.
  • AI Architecture: Multi-tier AI integration powered by Google Gemini (Gemma / Gemini 3.5 Flash) with automatic fallback to local OpenAI-compatible models, strictly grounded against lesson material to prevent hallucination.
  • Infrastructure & Integrations: Cloudflare Tunnels for zero-trust edge delivery, Cloudflare Email Workers for transactional auth emails, and Stripe for subscription management.

Challenges I ran into

  1. Zero-Defect Musical Generation: Music theory has strict mathematical and harmonic rules. Instead of hardcoding questions, we built a comprehensive music theory engine (theory.js) from scratch to calculate intervals, scale degrees, and clef transpositions programmatically.
  2. Pedagogical Adaptation: Crafting a prerequisite Directed Acyclic Graph (DAG) for learning objectives so the engine can gracefully backtrack and remediate root gaps when a student fails a mastery checkpoint.
  3. Safe, Grounded AI Tutoring for Learners: Preventing LLM hallucinations in educational contexts by strictly scoping prompts to syllabus objectives and approved lesson transcripts.

Accomplishments that I am proud of

  • Validated Question Bank: Validated over 14,000 algorithmic question generations with 0 defects.
  • Ultra-Fast Zero-Dependency Stack: Built a robust, full-featured web app, LMS, and AI proxy with zero bulky client-side frameworks or heavy npm bloat.
  • Full Learning Loop: Successfully connecting diagnostic quizzes, prerequisite remediation, sheet music engraving, synthesized audio playback, and contextual AI chat into one cohesive daily study habit.

What I learned

  • How cognitive models (like spaced repetition decay and mastery thresholds) can be directly translated into real-time adaptive code.
  • How to orchestrate resilient LLM fallbacks and bound concurrency to keep inference reliable and affordable.

What's next for SparcLearn

  • Expanding beyond music to all school curriculum
  • Adding teacher/classroom dashboards for multi-student cohort assignments.
  • Integrating Bayesian Knowledge Tracing (BKT) as more learning interaction data is collected.

Built With

Share this project:

Updates