Relevant Courses: Data Structures, Algorithms, Operating Systems, Number Theory, Prob & Stat I, Prob & Stat II, Real Analysis, Linear Algebra, Differential Equations

• Built MPI communication support for distributed LLM training and vLLM inference, enabling interchangeable MPI/NCCL backends across DDP, tensor, pipeline, and context parallelism on multi-GPU and multi-node HPC systems.
• Redesigned PyTorch's GPU-aware ProcessGroupMPI path, replacing device-wide GPU synchronization with CUDA event-based asynchronous execution, preserving compute-communication overlap and reducing training iteration time by 10%.
• Integrated MPI collectives into vLLM's compiled inference path, developing CUDA-stream-aware execution, stable-buffer handling, and piecewise CUDA Graph execution for tensor-parallel inference.
• Profiled and optimized MPI/NCCL communication using Nsight Systems and PyTorch Profiler, identifying synchronization, memory-registration, and GPU RDMA bottlenecks and improving 1 GiB AllReduce throughput from 49 to 176 Gbps.

• Built an autonomous multi-agent research system for NOAA (Python, MCP) that dynamically spawns parallel sub-agents with custom skills/tools to research topics, deployed in isolated Daytona sandboxes.
• Designed a NOAA organizational knowledge graph (257 nodes, 328 edges) with web-based visualization.
• Developed a Data Reach Metric algorithm aggregating multi-source signals (GitHub, policy, news) to quantify scientific ROI, giving stakeholders automated, narrative-driven funding justification reports.

• Trained multi-billion-parameter FORGE models (1.44B, 13B, 25.6B parameters, GPT-NeoX architecture) on a 257B-token dataset using the Frontier supercomputer, applying HPC-optimized GPU kernels such as FlashAttention tuned for AMD MI250X GPUs.
• Developed Python/Bash scripts and LMod modules to capture training metrics (loss, TFLOPS/sec, iteration time, learning rate, samples/sec) and integrated the OLCF-6 Acceptance Benchmark (FORGE) into Frontier's Test Harness, enabling automated validation of large-scale training runs against baseline performance thresholds to confirm system readiness for AI workloads.
• Compiled and integrated the AWS-OFI-RCCL plugin against Cray libfabric and ROCm to enable RCCL collective communication over Frontier's Slingshot interconnect; performed RCCL benchmarking across up to 1,024 AMD MI250X GPUs to evaluate communication patterns and identify training bottlenecks.

• Developed a full-stack conference management web app (NextJS, Firebase, Vercel) supporting 350+ registered users, with authentication, registration, venue/session lookup, and real-time push notifications for conference updates.

• Trained deep learning models on controlled subsets of the DeepFish dataset ranging from 1,000 to 40,000 images, achieving 99.95% classification accuracy across fish presence and absence in complex underwater environments.
• Developing a mathematical framework using nonlinear regression to model learning curves and estimate data requirements for deep-sea fish classification, localization, and segmentation tasks.

• Implemented and optimized global meshless methods (Kansa and MAPS) in MATLAB to approximate PDE eigenvalues on regular and irregular domains, enhancing numerical stability and reducing error by using ghost point parameters.
• Presented a comparative analysis on the impact of these methods at two academic conferences: the USM Undergraduate Symposium and the USA/USM/SELU Research Mini-Conference.

• Built an AI-powered web extension using LangChain and Microsoft Graph API to generate contextual email drafts and automate responses.
• Deployed backend on EC2 with a reverse proxy to handle server and AI model routes, enabling real-time draft generation.
• Added analytics to track user interactions and identify email trends for improved insights.

• Developed a freelancing platform for college students to offer small gigs (e.g., tutoring, hair cutting) using React for the frontend and Express.js on an EC2 server for the backend.
• Integrated GPT-4 API to generate tailored promotional content for student advertisements, enhancing user engagement.

• Built a mobile app with React Native (frontend) and Django (backend), deploying a pre-trained Kaggle model for plant disease detection to identify diseases in crops.
• Enhanced user experience with real-time weather alerts and push notifications for extreme weather conditions.

Presented research on meshless methods for approximating eigenvalues and eigenmodes of partial differential equations.

Participated in mathematical problem-solving contests and attended research talks in pure and applied mathematics.

Presented research on meshless methods for approximating eigenvalues and eigenmodes of partial differential equations.

Created a statistical model to forecast medal distributions at the 2028 Olympics; submitted findings in a paper.

Recieved funding from Cross Undergraduate Research Program to fund my research.

C++, Python, Node.js, JavaScript ES6+, Django, React, TensorFlow, PyTorch, Next.js, MATLAB, R

AWS, Google Cloud, PostgreSQL, Firebase, HPC, Slurm, RStudio