Redmond, Washington, United States
- Development of tool for increasing production order issue debugging efficiency Creating secure GraphQL endpoints that retrieve from an order’s ID order details as well as corresponding purchase transaction and payment request data. Developing front-end webpage for displaying an entire order’s details and lifecycle clearly, including error handling pathways for synchronous actions, asynchronous actions, and inconsistency job failures. Webpage also provides the ability to rerun asynchronous actions for a particular order, as well as provides preloaded links to other internal debugging pages for further debugging related information.
Formalizing a methodology to classify multimodal datasets based on the strength of their label dependence on individual modalities as well as the label's dependence strength on the interaction between modalities. Demonstrated that for an dataset with strong intra-modal dependence on a single modality, strong performance improvements can be seen with better modality-specific encoders or even strong unimodal models. Tested with the Cambrian-1 family of multimodal models, along with the benchmarks MME, MMMU, Mathvista, ScienceQA, and more. https://github.com/vm2781/cambrian_fork/tree/main/eval
Published company-wide standard format of New Relic queries to track different service metric RPS Developed RPS peak-finder algorithm on queries so test configuration can be matched to performance. Integrated peak-finder into GitLab pipelines for immediate analysis after production tests are complete, and Slack for casual viewing of scaled events. Helped transfer all New Relic dashboards into Terraform code to store core queries in localized place.
Exploring to improve efficiency and scalability for open-domain QA deep learning models. Developing new pipeline to pull relevant passages from key-value database as needed rather than storing on RAM to make running and testing the model accessible to users with less computational resources.
Utilized machine learning to both detect and predict anomalies in Nordstrom's large internal infrastructure. Collaborated with a peer to develop and deploy micro-service architecture supporting the models, including front-end web view of model-detected anomalies for clients to view and correct for model retraining on the backend. Presented results to Nordstrom's entire Infrastructure and Security division.