Luminary

Usage of Antigravity; training our ML model

✨ Luminary — Illuminate your path to mastery. "The mind is not a vessel to be filled, but a fire to be kindled." — Plutarch

💡 Inspiration: The "Where Do I Start?" Problem Education today is abundant, but direction is scarce.

Every day, millions of learners open YouTube, Reddit, or Coursera and face the exact same paralysis: "Where do I even start?" The internet is overflowing with amazing tutorials, yet the average self-taught learner spends more time searching for what to learn than actually learning it.

That paradox inspired Luminary. We wanted to build something that felt less like a static course platform and more like a personal academic advisor 🎓. We envisioned an AI companion that knows your current background, understands your specific goals, and lights the exact path forward step-by-step.

The ancient Greeks had an ideal called Paideia 🏛️ — education holistic and tailored to the individual. If Aristotle had access to machine learning, Google Calendar, and a Supabase backend, he would have built Luminary. 🚀

⚙️ What it does Luminary is an AI-powered personalized learning platform that dynamically generates optimized educational roadmaps. 🗺️

Instead of giving you a generic "Learn Python" checklist, Luminary:

Understands You: Evaluates your current skills via a precision diagnostic engine. 📊 Draws the Map: Generates a dynamic, interactive roadmap that builds a bridge from exactly where you are to where you want to be. 🛣️ Curates the Content: Provides targeted, high-quality resources (YouTube videos, articles, and customized AI-generated presentations) for every single node on your path. 📚 Fits Your Life: Integrates directly with Google Calendar 📅 to schedule learning sessions right into your actual daily routine. 🛠️ How we built it: The Architecture Luminary is built on a modern, robust full-stack architecture divided into four intelligent layers:

The Diagnostic Engine 📋 When a user signs up, they take a brief diagnostic test that gauges their baseline skills. Combined with their hobbies, interests, and timeframe, this data seeds their initial user profile in our Supabase backend.
The ML Recommendation Pipeline 🧠 This is the brain of the app. As users complete coursework, our Python-based recommendation engine goes to work:

It reads the user's completed topics and runs them through a Sentence Transformer to understand their semantic meaning. It traverses a Dependency Graph to see what they are ready for next (e.g., you need Algebra before Calculus). It scores every possible next topic using a composite algorithm: $$ \text{Score}(T_i) = \alpha \cdot \text{Readiness}(T_i) + \beta \cdot \text{InterestMatch}(T_i) + \gamma \cdot \text{UrgencyScore}(T_i) $$

The result? The system mathematically calculates the absolute best next thing for you to learn. 🎯

The Interactive Roadmap 🗺️ We built an interactive UI that turns complex graph data into a beautiful, expandable tree. 🌳

Resource Windows: Clicking a topic opens curated YouTube snippets, articles, and beautiful AI-generated PowerPoints (built on the fly via the Gemini API + PptxGenJS). 🖥️ Relational Reasoning: The UI actually explains why it recommended a topic ("We suggested Neural Networks because you just finished Linear Algebra"). 🔗

Smart Calendar Integration 📆 Learning only happens if you make time for it. Users can drag and drop learning sessions inside the app, and a single click uses OAuth 2.0 to sync those sessions directly to their Google Calendar.

💻 Our Tech Stack Frontend: Next.js 14 + Tailwind CSS ⚛️ (Sleek Dark Theme with #46b533 accents!) Backend / DB: Supabase (PostgreSQL + Auth) 🗄️ AI Pipeline: Python ML service + Gemini API 🤖 Presentations: PptxGenJS 📊 Calendar: Google Calendar API + NextAuth.js 📅 🚧 Challenges we ran into 🥶 The "Cold Start" Problem How do you recommend the perfect course to a brand-new user who has no history on your app? Relying strictly on their stated interests was too broad. We solved this by building a hybrid engine: we use content-based filtering for brand-new users, and smoothly transition to collaborative filtering (looking at similar users) as their history grows:

$$ \hat{r}{u,t} = \lambda \cdot \hat{r}^{\text{content}}{u,t} + (1 - \lambda) \cdot \hat{r}^{\text{collab}}_{u,t} $$

(Where λ decays as the user completes more content). 📈

🌳 Roadmap Rendering at Scale A roadmap for "Computer Science" has hundreds of deeply nested subtopics. Trying to render those all at once crashed the browser. We had to heavily optimize our React layer, using useMemo, dynamic imports, and virtualizing the tree to lazy-load nodes only when a user expands them. This kept the UI running at a buttery-smooth 60fps. ⚡

🤖 Taming the AI (Gemini outputs) Generating PowerPoint presentations on the fly using Gemini was amazing... when it worked. Initially, the AI's output varied wildly in structure, breaking our slide generator. We fixed this by enforcing a strict JSON-mode prompt schema and writing middleware to validate the AI's response before it ever touched the presentation generator. 🎯

🏆 Accomplishments that we're proud of Building a functional Machine Learning recommendation engine from scratch during a 24-hour time crunch! 🧠 Seamlessly integrating three massive external APIs (Gemini, Google Calendar, Supabase) into one snappy, cohesive interface. 🔌 Creating a beautiful, highly interactive frontend that turns intimidating, complex subject graphs into an inviting, playable roadmap. 🎨 Engineering AI-generated slide decks on the fly. It's one thing to link to a Wikipedia article; it's another to have an AI generate a custom masterclass presentation for you in seconds. 📽️ 📚 What we learned We didn't just build a learning app—we got an education ourselves:

Adaptive Learning Systems: We learned how to model skill progression mathematically using dependency graphs and readiness thresholds. Personalization at Scale: We realized that personalization isn't just filtering out bad content; it's heavily weighting the right content based on user urgency and interest embeddings. ⚖️ Full-Stack Resilience: Wiring a Next.js frontend to a Python ML backend, a Postgres DB, and external AI models taught us exactly how critical clean API boundaries and error handling are. 🏗️ 🚀 What's next for Luminary Luminary is just taking its first steps. Our future roadmap includes:

Peer Learning Groups — Connecting users who are currently stuck on the exact same roadmap nodes so they can study together. Spaced Repetition Integration — Automatically resurfacing older topics on your calendar right before you forget them, using algorithms like SM-2. Mobile App — Because learning shouldn't be chained to a desk! Instructor Mode — Enabling teachers and creators to build, customize, and share curated roadmaps with their audiences.