SmartClass – AI‑Powered Roadmaps & YouTube Learning for Students

SmartClass helps students turn outcomes (e.g., “learn Full‑Stack”) into actionable roadmaps, pairing each step with high‑quality YouTube lectures and practice tasks.

Project Name and Context

• Project: SmartClass (learning assistant)
• Audience: Students and self‑learners targeting practical outcomes
• Why: Abundance of content, shortage of structured paths

Functionality and Purpose

• Roadmap generation from a goal and skill level
• Curated lectures with quality filters (duration, engagement, recency)
• Practice tasks per node with difficulty ramp
• Progress tracking and reminders
• Offline notes and flashcards

Problem It Solves

• Reduces overwhelm; replaces random tutorials with structured tracks
• Finds current, credible videos that actually teach
• Encourages consistency via reminders + visible progress

Tech Stack Used

• Frontend: Next.js 15, Tailwind CSS, Zustand
• Backend: Node.js (API), PostgreSQL (progress), Redis (queues/cache)
• AI/RAG: OpenAI/LLM, vector store (pgvector/Weaviate), rerankers
• Ingestion: YouTube Data API v3, periodic refresh

How I Built It (or Am Building It)

Ingestion & Indexing

// fetch videos by topic, apply quality filters, store embeddings
for (const topic of topics) {
  const videos = await yt.search(topic, { max: 50 });
  const curated = videos.filter((v) => v.duration < 60 && v.views > 5000);
  await embedAndUpsert(
    curated.map((v) => ({
      id: v.id,
      title: v.title,
      desc: v.description,
    }))
  );
}

Roadmap Synthesis

// prompt + tools -> DAG of concepts with prerequisites
const roadmap = await llm.plan({ goal: "Full‑Stack Web", level: "beginner" });
// normalize to nodes/edges, attach evaluation rubrics

Retrieval & Ranking

// hybrid search: dense + lexical + rerank
const candidates = await searchVideos(node.topic);
const reranked = await rerank(candidates, node.learningOutcome);
return pickTop(reranked, 3);

Progress Tracking

CREATE TABLE progress (
  user_id UUID,
  node_id TEXT,
  status TEXT CHECK (status IN ('todo','doing','done')),
  updated_at TIMESTAMPTZ DEFAULT now(),
  PRIMARY KEY (user_id, node_id)
);

Unique Implementations or System Design

• DAG‑based roadmap with prerequisite checks
• Hybrid retrieval (dense + BM25) then learning‑outcome rerank
• Cold‑start filters: duration, engagement, recency, channel trust score
• Practice rubric to validate completion objectively

Realistic Challenges or Tradeoffs

• Ranking drift if you don’t refresh embeddings and signals
• YouTube API quotas; batch and cache aggressively
• Subjectivity in “quality” → use explicit signals + user feedback
• Motivation: build gentle nudges, not nagging

What I Learned or Improved

• Treat AI as an orchestrator; keep data models explicit
• Clear UX beats clever prompts—visibility drives completion
• Reliability: background jobs, retries, and observability from day one