Developed the “in the wild” fashion visual similarity model powering auto-tagging. Pioneered deploying both this and a low-latency object detection model together to be run on device using Apple’s CoreML framework. This model achieves 85% match rate in real-world usage. Built a personalized brand recommendations model for creators that, in a multi-week A/B test, outperformed the existing non-personalized solution with a 68.56% increase in clickthrough rate to the brand's website (p<0.0001). In addition, I've done a bunch of smaller things here, including deploying a search query refinement model, developing a solution for automatic brand categorization, hacking together a smart gift guide creator, working on a TON of POCs for future projects, and just a whole lot more.

Responsible for the organization's machine learning models, data analysis, and cloud infrastructure. Built and now maintain the Rep Score Portal to upload, track, and analyze bias in various content forms.

Trained and handed-off a production-ready transformer-based document classification model, DOOM, to better understand the content in articles across all the Condé Nast brands. This model outperformed the existing text-to-vector model, HAL, by over 35% in offline content recommendation tasks for a collection of articles and recipes across multiple different languages. Additionally improved monitoring for our newsletter recommendations jobs, reducing the job fail rate to <1%.

Built a cold-start optimized recommendations system for our talent-job matching algorithm and deployed the model in a custom-built API backed by Elasticsearch. This model resulted in a CTR increase of 55% in an email A/B test compared to the legacy matching model.Created a library and API for detecting fuzzy duplicate projects with low (< 50ms) latency.

Productionalized and open-sourced Collie, a novel deep learning recommendations library built with both flexibility and scalability in mind. In an email A/B test, this new member-product recommendations system showed a 184% improvement in click to open rate compared to our previously deployed recommendations system. Previously built a low-latency (<50ms), unsupervised fraud model for real-time use in detecting anomalous logins, pioneering unsupervised learning at ShopRunner via a novel evaluation metric called “cross scoring.” I also made a real-time, supervised fraud model for credit card transactions data, capable of reducing chargebacks by up to 52% more than the original rules-based system in place.

Built the initial framework for a deep learning recommendations library to outperform our original collaborative filtering model.

I worked to deprecate the current GD-IQ with a novel multi-task object recognition pipeline to identify, classify, and cluster character faces in media on perceived identities of gender, skin tone, age, and body type. These automated findings helped identify disparities in media with human-level annotation accuracy. As a data science team of one, I also found myself on a number of random smaller projects, including managing the Institutes’s AWS cloud infrastructure, building spaCy NER and fine-tuned BERT text classification models to automatically annotate data containing harmful language, deploying image and text annotation tools to be used in production by our research team, and roadmapping all machine learning projects to be completed by teams working at the Institute.

Collected, cleaned, and analyzed datasets to identify bias in media. Was responsible for running and maintaining the GD-IQ software to automatically analyze gender and race statistics in media.

Responsible for running lab sessions, grading assignments and exams, and holding office hours for students seeking additional information for both CS 331 (Discrete Structures and Algorithms) and CS 350 (Computer Organization and Assembly).

Collected, processed, cleaned, and ran the data analysis for a novel study examining stress and its effects in the workplace. Analysis included data visualization of stress signals and normalized stress level differences in box plots, natural language processing to determine sentiment and token type of subject responses, hypothesis testing and ANOVA to determine significant differences in physiological measures, and mixed linear models to explain what factors explain changes in stress levels.