Fahad
Hire me
← /projectsCase study · Capstone, AI/ML internship at Zynvex Solutions (remote, 6 weeks) · 2026
Zynvex Solutions logoZynvex Solutions

LearnTube — AI YouTube Learning Assistant

A Gemini-powered RAG application that turns any YouTube video into a summary, quiz, flashcards, and a transcript-grounded chat with [MM:SS] citations — shipped on web and native mobile.

Solo AI engineer — RAG pipeline, FastAPI backend, web + mobile clients

Intent classifier accuracy
~85%
Learning surfaces
Chat · Summary · Quiz · Cards
Clients shipped
Web + iOS/Android
Answer grounding
[MM:SS] citations
Python 3.13FastAPIPydanticGoogle Gemini FlashChromaDBsentence-transformersHuggingFaceONNX RuntimeTF-IDFLogistic Regressionscikit-learnSQLiteSQLAlchemyslowapiJWTReact 18ViteTailwind CSSshadcn/uiFramer MotionTanStack QueryExpo / React Native
Demo & preview

See it running.

A recorded product walkthrough plus screenshots taken from the project report.

View LearnTube project report (PDF) ↗Download report
Landing surface — paste any YouTube URL to start a study session.
Landing surface — paste any YouTube URL to start a study session.
Video workspace with the recent-videos library in the sidebar.
Video workspace with the recent-videos library in the sidebar.
Transcript-grounded chat answering with [MM:SS] citations.
Transcript-grounded chat answering with [MM:SS] citations.
Auto-generated summary: overview, key points, and chapters.
Auto-generated summary: overview, key points, and chapters.
Quiz generation with scored multiple-choice questions.
Quiz generation with scored multiple-choice questions.
Flashcard deck with reveal + mark-as-known progress.
Flashcard deck with reveal + mark-as-known progress.
Problem

Long-form YouTube lectures are a poor study medium: there is no way to ask a question, no reliable structure, and no way to self-test. Generic chatbots hallucinate because they never see the actual video content.

  • ▸Answer questions about a specific video using only that video's transcript.
  • ▸Generate study artefacts — overview, key points, chapters, quiz, flashcards — automatically.
  • ▸Stay frictionless: no sign-up required on the web, and full feature parity on mobile.
Approach

How I built it.

01

Retrieval pipeline

A FastAPI ingestion service fetches transcripts (youtube-transcript-api with a yt-dlp fallback), chunks them semantically while preserving timestamps, embeds each chunk with sentence-transformers (all-MiniLM-L6-v2), and stores vectors in ChromaDB. Retrieved chunks are injected into Gemini Flash prompts so every answer can cite the moment it came from.

  • ▸Timestamp-preserving chunking enables [MM:SS] citations in chat answers.
  • ▸ChromaDB vector store with ONNX embeddings keeps retrieval fast and local.
  • ▸SQLite + SQLAlchemy persist video metadata, summaries, and the user library.
02

Classical ML inside a GenAI app

A TF-IDF + Logistic Regression intent classifier (~85% accuracy) routes each user query to the right skill — RAG chat, summary, quiz, or flashcards — so the LLM is called with a purpose-built prompt instead of one generic prompt.

  • ▸scikit-learn model trained on labelled query intents.
  • ▸Deterministic routing keeps token cost and latency predictable.
  • ▸slowapi rate limiting and JWT verification protect the public API.
03

Product surfaces

The web client is React 18 + Vite with Tailwind CSS v4, shadcn/ui, Framer Motion, and TanStack Query; the mobile app is Expo / React Native with Expo Router and shared UI patterns. Both share the same REST API, and the web app supports anonymous browser-local identity for instant access.

  • ▸Chat, summary, quiz, and flashcard panels in a single video workspace.
  • ▸Browser-local library to revisit previous learning sessions.
  • ▸Packaged for autoscale deployment; walkthrough captured in the demo video and report.
Results

LearnTube was completed end-to-end on web and as a native mobile app, with grounded answers and auto-generated study material for arbitrary YouTube videos. It became the capstone deliverable of the Zynvex Solutions AI/ML internship — the demo video and project report below document the full working product (no public hosted URL is live at the moment).

  • ▸End-to-end GenAI system: ingestion, embeddings, retrieval, prompting, and evaluation.
  • ▸Grounded responses with timestamp citations sharply reduce hallucination.
  • ▸Feature parity across web and mobile from a single FastAPI backend.
Learnings

What I'd carry forward.

  • →Retrieval quality — chunking strategy and metadata — matters more than prompt wording.
  • →A small classical model in front of an LLM buys accuracy, speed, and cost control.
  • →Anonymous-first access removes the biggest drop-off point for a learning tool.

More case studies

Browse other deep-dives or jump back to the full project list.

AI-Enhanced VPN Traffic Classifier →SmartHire AI — Job Portal →All projects →