🌉 Vidya Setu — AI-Powered Adaptive Learning Platform

Adaptive quiz engine · 10 Indian languages · RAG-powered tutoring · Accessibility-first

🚀 Quick Start · 🎮 Demo · ✨ Features · 🏗️ Architecture

---

📖 What is Vidya Setu?

Vidya Setu (Bridge of Knowledge) is an AI-powered learning platform that bridges gaps in education—empowering disabled and underserved learners with adaptive quizzes and a RAG-based tutor that understands their study material and responds in their language and learning style.

It combines a ML-powered adaptive quiz engine with a RAG-based personal tutor that reads your study material and answers questions in your language, adapted to your accessibility needs.

🏆 Built for the Databricks Hackathon 2026 — demonstrating real-world AI for social good in Indian education.

---

✨ Features

Feature	Description
🧠 Adaptive Quiz Engine	FAISS + Sentence Transformers select questions by difficulty and weak topics
📈 Real-time Difficulty Adjustment	Automatically upgrades Easy → Medium → Hard based on performance
🌍 10 Indian Languages	English, Hindi, Tamil, Telugu, Bengali, Marathi, Gujarati, Kannada, Malayalam, Punjabi
🔊 Text-to-Speech	Sarvam AI reads questions and answers aloud in your chosen language
📄 RAG-Based PDF Tutor	Upload any textbook/notes → ask questions → get cited answers
♿ Accessibility Profiles	Adapted outputs for ADHD, Dyslexia, Visual Impairment, Hearing Impairment
📊 Live Analytics Sidebar	Topic mastery, difficulty breakdown, accuracy, and focus recommendations
🎯 Source Citations	Every RAG answer cites the exact page number from your PDF

---

🏗️ Architecture

flowchart TD
    User([👤 Student]) --> App[🖥️ Streamlit App]

    App --> QuizEngine[🧠 Adaptive Quiz Engine]
    App --> RAGEngine[📄 RAG Learning Engine]
    App --> TTS[🔊 Sarvam AI TTS]

    subgraph QuizEngine[🧠 Adaptive Quiz Engine]
        direction LR
        ST[Sentence Transformers\nall-mpnet-base-v2] --> FAISS[FAISS\nVector Index]
        FAISS --> Retriever[Smart Question\nRetriever]
        Retriever --> Difficulty[Difficulty\nPredictor]
    end

    subgraph RAGEngine[📄 RAG Learning Engine]
        direction LR
        PDF[PDF Upload\npypdf] --> Chunker[Text Chunker]
        Chunker --> BGE[BGE-small\nEmbeddings]
        BGE --> Similarity[Cosine\nSimilarity Search]
        Similarity --> LLM[LLaMA 3.3 70B\nDatabricks Serving]
    end

    subgraph Translation[🌍 Multilingual Layer]
        M2M100[M2M100 418M\nFacebook]
    end

    App --> Translation
    LLM --> App
    TTS --> App

Loading

---

🛠️ Tech Stack

🤖 AI / ML

Component	Technology
Large Language Model	LLaMA 3.3 70B via Databricks Model Serving
Quiz Embeddings	sentence-transformers/all-mpnet-base-v2
RAG Embeddings	BAAI/bge-small-en-v1.5
Vector Search	FAISS (IndexFlatIP)
Translation	Facebook M2M100 418M
Text-to-Speech	Sarvam AI Bulbul v3

🏗️ Platform & Infrastructure

Component	Technology
Deployment	Databricks Apps
Model Serving	Databricks Model Serving Endpoints
Frontend	Streamlit
PDF Processing	pypdf

---

🚀 Setup & Run

Prerequisites

• Python 3.9+
• Databricks workspace with Model Serving enabled
• Sarvam AI API key (get one here)

1. Clone & Install

git clone https://github.com/abhishek130904/DatabricksHackathon.git
cd DatabricksHackathon
pip install -r requirements.txt

2. Configure API Keys

Edit app.py and set your credentials:

# Databricks (auto-configured inside Databricks Apps)
DATABRICKS_HOST  = "https://<your-workspace>.azuredatabricks.net"
DATABRICKS_TOKEN = "your_databricks_token"

# Sarvam AI
SARVAM_API_KEY = "your_sarvam_api_key"

💡 Inside Databricks Apps, authentication is handled automatically — no token needed.

3. Run Locally

streamlit run app.py

App opens at http://localhost:8501

4. Deploy on Databricks Apps

# Import project to Databricks workspace
databricks workspace import-dir . /Workspace/Users/<your-email>/vidya-setu

Then in the Databricks UI:

1. Navigate to Databricks → Apps
2. Click Create App
3. Select /Workspace/Users/<your-email>/vidya-setu
4. Set environment variables (SARVAM_API_KEY)
5. Click Deploy — allow 8–10 minutes for large ML models to load

---

🎮 Demo Walkthrough

🧠 Demo 1: Adaptive Quiz Engine

1. Open app → Select subject (Physics / Math / Chemistry)
2. Answer 3 easy questions correctly
3. Watch difficulty auto-upgrade: EASY → MEDIUM
4. Switch language to Hindi using the Language dropdown
5. Click 🔊 "Read Aloud" — hear the question in Hindi
6. Check sidebar → see Topic Mastery and Weak Topic focus update live

📄 Demo 2: RAG-Powered Personalized Learning

1. Click "Personalized Learning" tab
2. Upload any PDF (textbook chapter, lecture notes, etc.)
3. Select Accessibility Profile → "ADHD"
4. Ask: "What is the main concept in this document?"
5. Observe: answer broken into numbered steps (ADHD adaptation)
6. Click 🔊 "Listen" — audio plays with duration shown
7. Expand 📚 Sources → see exact page numbers cited
8. Switch profile to "Dyslexia" → ask again → notice simpler language

⚡ Demo 3: Quick Prompts to Impress Judges

"Explain Newton's Laws in simple terms"
"What are the key formulas mentioned in this chapter?"
"Summarize the important points from page 3"
"Give me a step-by-step explanation of this concept"

---

📁 Project Structure

vidya-setu/
├── app.py                 # Main Streamlit application
├── pcmDataset.csv         # Physics/Chemistry/Math quiz dataset
├── requirements.txt       # Python dependencies
├── app.yaml               # Databricks Apps configuration
└── README.md

---

🐛 Troubleshooting

Issue	Fix
LLM endpoint not found	Run pip install --upgrade databricks-sdk and verify endpoint name
TTS returns 0:00 duration	Raw PCM is auto-wrapped in WAV header — ensure speech_sample_rate: 8000 in payload
Translation model slow	M2M100 loads once and is cached — wait for first load (~2 min)
Slow startup (8–10 min)	Normal — large ML models (M2M100, Sentence Transformers) load on cold start
PDF embeddings fail	Ensure pypdf and sentence-transformers are installed; try a text-based PDF
FAISS import error	Run pip install faiss-cpu (or faiss-gpu on GPU instances)

---

📦 requirements.txt

streamlit>=1.32.0
pandas
numpy
sentence-transformers
faiss-cpu
scikit-learn
transformers
torch
pypdf
databricks-sdk
requests

---

♿ Accessibility Profiles

Profile	Adaptation Strategy
🔵 Default	Balanced explanation with examples
👁️ Visual Impairment	Rich text descriptions, no visual references, full verbal context
👂 Hearing Impairment	Complete written output, no audio-dependent phrasing
📖 Dyslexia	Short sentences, simple vocabulary, bullet points
🧠 ADHD	Numbered steps, concise chunks, high-focus structure

---

🌍 Supported Languages

Language	Native Script	TTS Support
English	English	✅
Hindi	हिंदी	✅
Tamil	தமிழ்	✅
Telugu	తెలుగు	✅
Bengali	বাংলা	✅
Marathi	मराठी	✅
Gujarati	ગુજરાતી	✅
Kannada	ಕನ್ನಡ	✅
Malayalam	മലയാളം	✅
Punjabi	ਪੰਜਾਬੀ	✅

---

👨‍💻 Contributors

Abhishek Raj IIT Indore @abhishek130904

Purvi Jain IIT Indore

Adarsh Rai IIT Indore

Lakshya Rishi IIT Indore

---

---

📄 License

Open-source for educational purposes. Built for Databricks Hackathon 2026.

---

If Vidya Setu helped you, please ⭐ star the repo!

Bridging knowledge gaps with AI — one student at a time.