🧠PrepSaarthi (Multimodal AI Companion)

An Autonomous Multimodal AI Learning Companion featuring distraction-free video processing, Indic language dubbing, and mathematically rigorous exam proctoring.

---

🌟 The Vision

In a world of short attention spans and passive consumption, we built a platform that forces active learning. This platform takes standard educational content (like YouTube videos) and transforms it into a highly structured, distraction-free, and accessible learning environment. Furthermore, it tests that knowledge using an uncompromising, AI-powered proctoring engine.

✨ Key Features

🎬 1. Multimodal Content Processor (The Smart Player)

• Distraction-Free UX: Strips YouTube videos of comments, recommendations, and algorithmic traps using custom iframe parameters.
• Instant Knowledge Extraction: Bypasses manual note-taking by utilizing youtube-transcript-api to instantly extract video text.
• Groq Llama-3 AI Engine: Generates highly structured Markdown notes and interactive Mermaid.js mindmaps from the transcript at 800+ tokens per second.
• On-the-Fly Indic Dubbing: Translates the entire lecture to Hindi using Groq, synthesizes it into audio using gTTS, and perfectly synchronizes playback with the user's video controls.

👁️ 2. The Uncompromising Proctor (Zero-Trust Exam Engine)

• Head-Pose Tracking: Utilizes @vladmandic/face-api (TinyFaceDetector) running entirely client-side to detect if a student drops their gaze to look at a phone.
• Environment Locking: Native Browser APIs strictly enforce Fullscreen mode and track Page Visibility to instantly catch tab-switching or Googling attempts.
• Merciless Flagging: A built-in tolerance window catches micro-flickers, issuing UI warnings and automatically submitting the exam upon repeated violations.

🧠 3. The SRS Data Loop (Spaced Repetition)

• Mistakes made during proctored mock tests are not forgotten. They are logged into a MongoDB Vault and rescheduled based on the Ebbinghaus Forgetting Curve to optimize daily revisions.

---

🏗️ System Architecture

We utilized a robust microservice approach to ensure fault tolerance during the live demo:

• Frontend: React + Vite + Tailwind CSS.
• Backend: Python FastAPI for heavy asynchronous AI processing and data routing.
• Database: MongoDB (Motor Async Driver).
• AI Routing: * Groq API (Llama 3 70B): Dedicated exclusively to the heavy Multimodal Content Processor for blazing-fast text generation and translation.
  • Google Gemini 2.5 Flash: Powers the dynamic Quiz Generation and Vault summarization.

---

🚀 Local Installation & Setup

Prerequisites

• Node.js (v18+)
• Python (3.10+)
• MongoDB (Local or Atlas URI)

1. Clone the Repository

```bash git clone https://github.com/yourusername/your-repo-name.git cd your-repo-name ```

2. Backend Setup (FastAPI)

Open a terminal and navigate to the backend folder: ```bash cd backend

Create and activate a virtual environment

python -m venv venv source venv/bin/activate # On Windows: venv\Scripts\activate

Install dependencies

pip install fastapi uvicorn groq google-generativeai youtube-transcript-api gTTS motor pydantic

Set up environment variables

cp .env.example .env ``` Configure your .env file: ```env MONGO_URI=mongodb://localhost:27017 GROQ_API_KEY=your_groq_key_here GEMINI_API_KEY=your_gemini_key_here ``` Start the Server: ```bash uvicorn main:app --reload --port 8000 ```

3. Frontend Setup (React/Vite)

Open a terminal and navigate to the frontend folder: ```bash cd frontend

Install dependencies

npm install

Start the development server

npm run dev ```

4. Face-API Models Configuration

Ensure the TinyFaceDetector neural network weights are present. Download tiny_face_detector_model-weights_manifest.json and tiny_face_detector_model.weights.bin and place them inside the frontend's public/models directory.

---

🛠️ Challenges We Ran Into (Hackathon Realities)

• The WASM Bundler War: We initially attempted to build our proctoring engine using WebGazer.js. However, modern Vite bundlers clashed violently with Google's MediaPipe WASM binaries, throwing continuous 404 errors. We engineered a hard pivot to @vladmandic/face-api, trading exact pupil-tracking for highly accurate head-pose (phone-checking) tracking, which integrated flawlessly with React.
• Audio-Video Synchronization: Making an AI-generated .mp3 file pause, play, and seek perfectly in sync with a third-party YouTube Iframe required writing custom React useRef hooks to forcefully tether the HTML5 audio element's clock to the Iframe's emitted progress events.

---

👥 Team

Built with late nights and a lot of caffeine by Team kaala khatta.

• Rahul Sahu
• Shourya Sinha
• Ashutosh Behera

---

Note: The backend is currently configured to use an instant-text testing route via youtube-transcript-api to avoid live stage-demo timeout risks associated with downloading raw YouTube audio.