Hi, I'm Yeswanth Chirumamilla

Project Overview

A comprehensive web scraping toolkit built with Python and Django, featuring multiple specialized modules for different data extraction needs.

Key Features

• Movie Data Aggregator: Scrapes movie information from multiple sources including IMDb, Rotten Tomatoes, and other film databases
• Weather Information System: Real-time weather data extraction from various meteorological websites
• Price Comparison Tool: Automated price monitoring across e-commerce platforms
• Intelligent Doubt Resolver: Uses BERT model to summarize answers from 5+ educational websites

Technical Implementation

• Built with Python using Selenium for dynamic content scraping
• BeautifulSoup for HTML parsing and data extraction
• BERT transformer model for natural language processing
• Django framework for web interface and API endpoints
• Implemented rate limiting and ethical scraping practices

Challenges Solved

• Handling dynamic JavaScript-rendered content
• Implementing robust error handling for network failures
• Creating efficient data storage and retrieval systems
• Ensuring compliance with website terms of service

Project Overview

A comprehensive comparative study of machine learning approaches for movie sentiment analysis, demonstrating the effectiveness of traditional ML vs deep learning methods.

Model Performance

• SVM Model: Achieved 86.9% accuracy with optimized hyperparameters
• RNN (LSTM) Model: Achieved 50% accuracy, highlighting overfitting challenges
• Key Finding: Traditional ML outperformed deep learning on this specific dataset

Technical Implementation

• Data preprocessing with NLTK for text cleaning and tokenization
• Feature extraction using TF-IDF vectorization
• SVM implementation with scikit-learn and hyperparameter tuning
• LSTM network built with TensorFlow/Keras
• Cross-validation and performance metrics analysis

Dataset and Methodology

• Used IMDb movie reviews dataset with 50,000 reviews
• Implemented stratified train-test split (80-20)
• Applied text preprocessing: lowercasing, punctuation removal, stopword filtering
• Compared multiple evaluation metrics: accuracy, precision, recall, F1-score

Key Insights

• SVM's linear decision boundary worked well for sentiment classification
• LSTM suffered from overfitting due to limited training data
• Feature engineering proved more important than model complexity

Project Overview

A full-featured real-time chat application built with the MERN stack, providing seamless communication with modern web technologies.

Core Features

• Real-time Messaging: Instant message delivery using Socket.io
• User Authentication: Secure login/signup with JWT tokens
• Friend System: Send and accept friend requests
• Online Presence: Real-time user status indicators
• Message History: Persistent chat storage and retrieval

Technical Architecture

• Frontend: React.js with hooks and context API for state management
• Backend: Node.js with Express.js RESTful API
• Database: MongoDB with Mongoose ODM
• Real-time: Socket.io for bidirectional communication
• Authentication: JWT tokens with bcrypt password hashing

Advanced Features

• Message encryption for security
• File sharing capabilities
• Emoji support and reactions
• Responsive design for mobile and desktop
• Message search and filtering

Deployment & Scalability

• Deployed on Render.com with automatic deployments
• MongoDB Atlas for cloud database hosting
• Implemented connection pooling for database efficiency
• Error handling and logging for production stability

Project Overview

Advanced reinforcement learning research project comparing online learning optimizers in the complex, dynamic environment of StarCraft II mini-games. Implemented and compared FTRL, β-FTRL, and PPO (Proximal Policy Optimization) algorithms for policy learning and optimization.

Research Findings

• β-FTRL Performance: Achieved 2x reward improvement compared to standard FTRL
• PPO Efficiency: PPO provided stable and efficient policy updates, outperforming DQN in this context
• Training Efficiency: β-FTRL and PPO reached optimal performance in less time
• Convergence Stability: More stable learning curves with reduced variance

Technical Implementation

• Environment: StarCraft II Learning Environment (SC2LE)
• Algorithms: FTRL, β-FTRL, and PPO optimization comparison
• Computer Vision: OpenCV for game state analysis
• Neural Networks: PPO-based policy networks for action selection
• Data Processing: Real-time game state feature extraction

Mini-Games Tested

• MoveToBeacon: Basic movement and targeting
• CollectMineralShards: Resource collection optimization
• FindAndDefeatZerglings: Combat strategy learning
• DefeatRoaches: Advanced unit micro-management

Key Contributions

• Demonstrated β-FTRL and PPO's superiority in dynamic environments
• Developed novel reward shaping techniques for SC2
• Created efficient state representation for real-time strategy games
• Published comparative analysis of online learning optimizers

Challenges Overcome

• Handling partial observability in game states
• Managing large action spaces efficiently
• Balancing exploration vs exploitation in dynamic environments
• Optimizing computational performance for real-time learning

Project Overview

An end-to-end spam classification system using Natural Language Processing and Artificial Neural Networks to accurately identify spam messages in SMS and email communications.

Model Performance

• Accuracy: Achieved 97.2% accuracy on test dataset
• Precision: 96.8% precision for spam detection
• Recall: 95.4% recall ensuring minimal false negatives
• F1-Score: 96.1% balanced performance metric

Data Preprocessing Pipeline

• Text Cleaning: Removal of special characters, URLs, and numbers
• Tokenization: Breaking text into individual words
• Stemming: Reducing words to their root forms using Porter Stemmer
• Stopword Removal: Filtering out common words with low predictive value
• Vectorization: Converting text to numerical features using Bag-of-Words

Neural Network Architecture

• Input Layer: 5000 features from TF-IDF vectorization
• Hidden Layers: 3 dense layers with ReLU activation (128, 64, 32 neurons)
• Dropout: 0.3 dropout rate to prevent overfitting
• Output Layer: Single neuron with sigmoid activation for binary classification
• Optimizer: Adam optimizer with learning rate scheduling

Dataset and Features

• SMS Spam Collection Dataset with 5,574 messages
• Balanced dataset with 4,825 ham and 747 spam messages
• Feature engineering including message length, word count, and character frequency
• Cross-validation with stratified k-fold for robust evaluation

Technical Implementation

• Libraries: TensorFlow/Keras for neural networks, NLTK for NLP
• Data Handling: Pandas for data manipulation, NumPy for numerical operations
• Visualization: Matplotlib and Seaborn for performance analysis
• Model Persistence: Saved trained model and vectorizer for deployment

Project Overview

This project explores and compares two different approaches to wake word detection in speech recognition systems. It evaluates a lightweight, real-time streaming model (Porcupine) and a more computationally intensive chunk-based speech-to-text model (Vosk). The goal was to measure their performance, efficiency, and suitability for edge device applications.

Key Features

• Porcupine (by Picovoice):
  • Real-time, low-latency wake word detection
  • Consumes <4% CPU and ~25 MB RAM
  • Ideal for resource-constrained environments like Raspberry Pi or mobile devices
• Vosk:
  • General-purpose speech-to-text engine
  • Operates on 2-second audio chunks
  • Consumes >20% CPU and ~180 MB RAM
  • More flexible but less efficient for continuous wake word detection

Technologies Used

• Python
• Picovoice Porcupine SDK
• Vosk API
• NumPy, SoundDevice, PyAudio
• CPU and memory benchmarking

Outcome

• Porcupine proved significantly more efficient for wake word detection on edge devices.
• Vosk showed better general speech transcription capabilities but is not optimized for real-time keyword spotting.