Machine learning software engineer. My academic research was in discrete signal processing and machine learning for brain-computer interface applications, with a focus on detecting imagined movements from brainwaves.
GitHub: https://github.com/pwstegman
Website: https://pwstegman.me
Designed and implemented a reinforcement learning model to adaptively rerank search results. This became a key marketing point for Yext's search product. Conducted research and prototyped using Python 3, Pandas, and NumPy. Productionized the final model in Java.
•
Designed a feed-forward neural network to estimate which search results a user is most likely to click. Used Pandas and NumPy to gather and clean historic user interaction data. Used TensorFlow to train the model to output a probability estimate for a user click when given the BERT vector for a query and a result.
•
Domain expert for GPT-3 initiatives at Yext. Trained a large GPT-3-like model for few-shot natural language processing (NLP) tasks. Vetted hardware vendors, determined appropriate cost/performance tradeoff, optimized existing libraries (Megatron-LM + DeepSpeed and GPT-NeoX) for our hardware, and researched ideal prompt design and prediction parameters.
•
Designed a Vue.js web app for visualizing BERT self-attention. Used Python and PyTorch methods to reduce the attention matrices down to a single summary matrix. Exposed Python methods with a Flask REST API. This demo was used to sell our search product.
•
Designed a set of ML model serving components, allowing for quicker turnaround times on model deployment. Oversaw the modularization of our monolithic ML serving application, ensuring we can easily scale our codebase as we grow.
•
Wrote department-wide guidelines for system architecture design, ensuring all code followed a set of standards for easier collaboration and better maintainability.
Developed an Angular application for the U.S. Dept. of Health and Human Services (HHS) to search through and filter contracts for health supplies using natural language. Users could rate filtered results as thumbs-up or down, which updated a reinforcement learning model.
•
Developed a data-labeling web app for machine learning research within HHS.
