vecml.com

Training large language models on named-entity recognition and text classification tasks. Also working with multi-modal models on time series problems.

Undergraduate researcher at the JSALT workshop. Produced large scale graph embeddings with traditional linear algebra methods, wavelets and Fourier transforms. Used truncated singular value decomposition and Chebyshev iterations.

Reimplemented and refined Transformer models using Pytorch and Neuron SDK for AWS EC2 instance. BEiT achieved 83.96% accuracy on ImageNet validation set, and ViT achieved 76.92%.Measured CPU compiled models against GPU runs, benchmarked throughput, latency, and cost.Created repository for SDEs/SAs on AWS-Neuron team which is extendible to future models.

Used Java, Maven, and Selenium to automate test suite, reduced test workload from 96 hours to 4. Automated for Optum Financial, which serves 1.8M customers who pay 200M in claims annually.Worked in agile development cycle, collaborated with development teams, management, PMs.

Intern in the IT department, worked with system installation, used Linux.