A Little About Me :
💻 Currently pursuing Master of Science in Computer Science at Rochester Institute of Technology
🔋 Bachelor of Technology in Electronics and Communication from Vellore Institute of Technology
⚙️ Experienced in building scalable platforms, robust backend systems, and AI-driven tools that solve real-world problems.
📊 I crunch numbers, chase data, and love figuring out how things tick.
🕹️ Gaming strategist extraordinaire, I dive into the depths of complex games to uncover winning tactics.
💸 A multi-agent, AI-powered trading simulation framework for backtesting and optimizing investment strategies on real financial data.
🤖 Implements LLM-driven trading agents that analyze market signals, historical data, and sentiment to automate adaptive portfolio decisions.
🔄 Supports flexible agent design, allowing comparison and rapid experimentation with custom trading strategies (e.g., value, momentum, sentiment-based).
📊 Features a robust backtesting engine to benchmark agent performance, visualize returns and drawdowns, and compare results against market indices.
🧠 Integrates financial sentiment analysis and news signals for context-aware, realistic trading simulations.
🛠️ Designed for extensibility, supporting new agent types, data sources, and analytics features for advanced research and innovation.
💻 A multi-functional AI-powered chatbot that interacts with diverse file types, including PDFs, Word documents, and text files.
📂 Extracts content from uploaded files, processes it into manageable chunks, and enables conversational interaction.
🚀 Features advanced Natural Language Processing (NLP) capabilities powered by OpenAI APIs for chat and summarization.
📊 Equipped with similarity search using Annoy, enabling precise query-based retrieval of information from documents.
🛠️ Designed for scalability and extensibility, supporting additional file types and integrations for future use cases.
🌐 Designed and built a SmartWebCrawler to automate the creation of machine learning datasets by extracting domain-specific data from the web.
🔍 Filters and collects structured content such as research articles, publications, and datasets using advanced HTML parsing with BeautifulSoup and automated browsing with Playwright.
⚙️ Supports custom crawling configurations, including URL filters, crawling depth, delays between requests, and output formats, making it adaptable for varied data requirements.
📊 Automatically preprocesses and structures crawled data for direct integration into machine learning pipelines.
🛡️ Handles content restrictions like paywalls and captchas for seamless data access, ensuring comprehensive dataset collection.
🚀 Scalable and extensible architecture, capable of supporting additional data types and integrations for building specialized ML datasets.
🌍 Built a federated learning framework using Flower and PyTorch, enabling decentralized model training across multiple clients.
📊 Implemented Trust & Reputation Mechanisms 🏆 to mitigate adversarial attacks, ensuring secure and reliable model aggregation.
🧪 Simulated data poisoning attacks by flipping labels in Clients 9 & 10, analyzing their impact on global model performance.
⚡ Designed custom aggregation strategies to enhance robustness against unreliable client updates.
📈 Achieved scalable and privacy-preserving training, enabling collaborative learning without sharing raw data.
🚀 Includes detailed visualization plots to track loss, accuracy, and the effectiveness of trust-based aggregation strategies.
📚 Developed an UDA framework to enhance text classification performance in low-resource settings by leveraging unlabeled data for improved generalization.
📊 Achieved a 40% reduction in error rates on datasets such as IMDb, Yelp, Amazon, and DBpedia by enriching training data with advanced augmentation techniques.
🔍 Utilized BERT transformers and back-translation to generate semantically diverse and robust training data.
⚙️ Designed a TF-IDF-based word replacement technique** to preserve critical keywords, optimizing topic classification performance.
🚀 Achieved 62.5% evaluation accuracy with minimal labeled data, showcasing UDA’s scalability and adaptability for real-world applications.
⭐ A high-performance Convolutional Recurrent Neural Network (CRNN) designed for wake word detection in voice-based applications.
🔊 Accurately detects wake words from audio streams, achieving an impressive 98% accuracy by leveraging advanced CRNN architectures.
📈 Designed for low-latency processing to ensure seamless integration with real-time systems.
🔍 Utilizes preprocessing techniques such as spectrogram generation to convert audio signals into model-friendly formats.
🚀 Built using TensorFlow and Keras, ensuring compatibility with modern deep learning frameworks.
⚙️ Supports transfer learning for domain-specific wake word customization, making it adaptable for various industries.
🛠️ Extensible and optimized for scalability, enabling deployment on edge devices and cloud platforms.
📊 Extracts and analyzes text from web pages and transcripts to identify sentiment for specific keywords.
🌐 Leverages BeautifulSoup for content extraction and supports transcript processing using speech_recognition.
🔍 Utilizes NLTK and WordNet for text tokenization, lemmatization, and synonym detection.
⚙️ Determines sentiment using TextBlob and presents results in a structured pandas DataFrame.
🚀 Configurable and extensible design for supporting additional keywords and content sources.