ASL Interpretation
Developed a model for continuous American Sign Language (ASL) interpretation using LLaVA-NeXT-Video, fine-tuned on How2Sign dataset with parameter-efficient techniques (LoRA and QLoRA), enhancing accessibility by translating ASL gestures into coherent English text
Python, PyTorch, LLVM, LLaVA-NeXT-Video, QLoRA
RAG Chatbot for Research Papers
Built RAG chatbot for querying research papers, enabling efficient retrieval and generation of context-aware insights to enhance accessibility; integrated LLM metrics tracking to optimize performance and cost, and deployed with a CI/CD pipeline for scalability
Python, Pinecone, LangChain, OpenAI GPT-3.5-Turbo LLM, Langtrace, Kubernetes
GPU Power Optimization using Frequency Scaling
Researched GPU energy optimization using clock frequency scaling strategies, leveraging Accel-Sim framework to analyze workload behaviors across NVIDIA architectures; achieved insights into energy-aware configurations for diverse computational tasks
Accel-Sim, AccelWattch, CUDA
ResNet-18 for EMNIST Classification
Implemented and trained a ResNet-18 model for classifying handwritten digits and uppercase letters using the EMNIST dataset. Evaluated model performance with precision, recall, and F1-score metrics, achieving 93.35% validation accuracy.
Python, TensorFlow, Keras, EMNIST, ResNet-18, Scikit-learn, Matplotlib
Object Matching with Hungarian Algorithm
Implemented bipartite matching of objects across surveillance frames using a custom cost matrix combining IoU and centroid distance; applied the Hungarian algorithm for optimal assignment and visualized results with color-coded bounding boxes
Python, OpenCV, NumPy, SciPy, Matplotlib, Hungarian Algorithm
Performance Analysis of Conv2d Layer in CNNs
Conducted a comprehensive performance analysis of the Conv2d-2 layer in a Convolutional Neural Network (CNN) trained on the MNIST dataset; evaluated both theoretical and empirical metrics, focusing on FLOPs (Floating Point Operations) and memory usage across different batch sizes on an NVIDIA V100 GPU using PyTorch
Python, PyTorch, NVIDIA Nsight Compute
House Value Prediction
Developed a machine learning model using XGBoost and other techniques to predict residential property prices in Ames, Iowa, achieving R² score of 0.918; provided actionable insights for real estate stakeholders to optimize investment strategies
Python, Scikit-learn, XGBoost, Pandas, Matplotlib, SHAP values
Multithreaded Run-Length Encoder
Designed and implemented a scalable thread pool in C for a Run-Length Encoding (RLE) utility, incorporating robust synchronization mechanisms to ensure thread safety, enabling efficient parallel processing and significantly optimizing application performance
C, Multithreading, Pthreads, Run-Length Encoding, Data Compression, Parallel Processing
Bookstore Database & Recommendation System
Designed and implemented a bookstore management system with integrated machine learning models to optimize inventory management and provide personalized customer recommendations, enhancing both operational efficiency and customer experience
Python, Scikit-learn, MySQL, GCP, Flask, Pandas, SQL
Hill Climbing Algorithm for Optimization Problems
Implemented a Python-based hill climbing algorithm to solve the N-Queens and Knapsack problems, integrating features like random restarts and sideways moves to optimize search efficiency.
Python, Algorithm Design, Combinatorial Optimization, Heuristic Search, N-Queens Problem, Knapsack Problem
File Recovery Tool
Developed a file recovery tool in C for FAT32 file systems, capable of recovering both contiguous and non-contiguous deleted files; implemented features such as SHA-1 validation for file integrity and optimized data retrieval using direct disk mapping
C, Unix, Operating Systems, FAT32 File Systems, SHA-1, Disk Mapping, System Programming
Markov Process Solver
Developed a Markov Decision Process (MDP) solver in Python, implementing value iteration and policy improvement algorithms to optimize decision-making for MDPs; supported configurable discount factors, tolerance levels, and iteration limits to accommodate various scenarios in reward maximization and cost minimization
Python, Markov Decision Processes, Value Iteration, Policy Improvement, Algorithm Design, Optimization
NFT Marketplace DApp
Developed a decentralized application (DApp) on the Ethereum blockchain for trading, minting, buying, selling, and auctioning NFTs, utilizing Solidity for smart contract development and React for a dynamic frontend. Implemented key features including batch minting, royalty systems, and secure trading mechanisms
Solidity, Ethereum, ERC-721, Web3.js, React, Node.js, Truffle, Ganache, MetaMask
AI-based Logic Solver
Developed an AI-driven logic solver in Python, integrating advanced algorithms for converting logical expressions into CNF and determining their satisfiability using the DPLL method. The system supporting complex decision-making and automated reasoning in AI applications
Python, Logical Reasoning, SAT Solvers, BNF to CNF Conversion, DPLL Algorithm, Algorithm Design
Analysis of Library Training Effectiveness
Analyzed survey data from 1,000 students to assess SMU Library's training effectiveness; identified significant improvements in research skills using statistical tests and visualizations, leading to actionable training recommendations
Tableau, JMP, Data Cleaning, Statistical Analysis, Hypothesis Testing, Data Visualization