About Me
About Me
I'm a graduate student at Northeastern University pursuing my MS in Information Systems, with a focus on machine learning and systems engineering. Before moving to Boston, I worked as a software engineer at Bosch in India, where I developed a deep appreciation for building reliable systems at scale. My interests lie at the intersection of cutting-edge ML research and practical engineering. I'm fascinated by the challenge of making sophisticated models accessible, whether that's optimizing memory footprints, improving inference speeds, or building intuitive interfaces. I believe the best technical solutions are invisible to their users, and I apply this philosophy whether I'm working with neural networks or traditional software systems.
I'm a graduate student at Northeastern University pursuing my MS in Information Systems, with a focus on machine learning and systems engineering. Before moving to Boston, I worked as a software engineer at Bosch in India, where I developed a deep appreciation for building reliable systems at scale. My interests lie at the intersection of cutting-edge ML research and practical engineering. I'm fascinated by the challenge of making sophisticated models accessible, whether that's optimizing memory footprints, improving inference speeds, or building intuitive interfaces. I believe the best technical solutions are invisible to their users, and I apply this philosophy whether I'm working with neural networks or traditional software systems.
Projects
A selection of my recent work and contributions.
Large-Scale Scene Recognition
Built scalable computer vision system for 365-class scene understanding using PyTorch on 1.8M images, achieving 20-30× speedup through multi-GPU distributed training on HPC cluster. Optimized deep learning pipeline across 1-4 GPU configurations, reducing training time while maintaining accuracy.
Formula 1 prediction system using ensemble machine learning models to forecast qualifying and race results. Features advanced feature engineering, real-time data integration via the FastF1 API, and sophisticated, F1-specific evaluation metrics.
A full-stack, multimodal Retrieval-Augmented Generation (RAG) application that allows users to upload PDF documents and engage in a real-time, conversational Q&A. The backend, built with FastAPI, handles PDF parsing, advanced semantic chunking, and a dual-strategy (semantic + keyword) retrieval from a ChromaDB vector store.
LightsTrail is a web app that provides real-time aurora forecasts based on your location, helping travelers, photographers, and enthusiasts catch the Northern Lights. Users can set alerts for aurora activity and share sightings, photos, and tips with a global community.
Chrome extension with 10K+ daily users featuring secure Web Crypto API encryption with AES-GCM and PBKDF2. Implements dual-mode architecture with proxy mode and BYOK functionality for ChatGPT/Claude prompt enhancement. Built with vanilla JavaScript and serverless Cloudflare Workers backend achieving sub-100ms response times.
Implemented QLoRA fine-tuning pipeline for 7B parameter LLMs using PyTorch and Hugging Face Transformers, achieving 75% VRAM reduction through 4-bit quantization. Optimized training with mixed-precision computation and gradient accumulation, enabling stable convergence on consumer GPUs with 15GB memory constraints.
Multi-agent system using LangGraph and Claude SDK with 5 specialized agents for automated code review, achieving 90% vulnerability detection accuracy. Implemented model routing with LangChain/OpenRouter selecting between Claude 3.5 and Gemini models, optimizing costs by 40%. FastAPI backend processes 10K line repositories in under 20 seconds.
Bayesian Analysis of International Trade Data
Automated data pipeline processing 30K+ records from UN/World Bank APIs, creating unified 15-year longitudinal dataset. Built PyMC3 Bayesian model with Monte Carlo simulation (10K+ iterations) quantifying trade policy impacts. Created predictive decision tool forecasting $0.4B-$625B economic impact scenarios for FIFA 2026.
Developed Python MCP server connecting AI agents to enterprise financial APIs, implementing SQL queries for real-time data retrieval and analysis. Created intelligent caching and validation framework ensuring data integrity for prediction tasks in production workflows.
Computer Vision for Video Understanding
Developed emotion recognition system for videos using CNN/LSTM architectures with PyTorch, processing temporal sequences for action recognition achieving 84% accuracy. Built image classification system implementing transfer learning with ResNet and VGG. Created video captioning model combining CNN encoders with LSTM decoders for multi-modal understanding.