Kilbourne Charrier

Language is ubiquitous to the modern online experience. In particular, there has been immense progress on large language models (LLMs) over the past few years: LLMs have demonstrated amazing capabilities across a broad range of tasks and have been referred to as "foundation models". However, LLMs can also be harmful to online conversations without proper guardrails in place as they can generate toxic texts, reflect human biases, as well as introduce its own biases in the training process. Left unchecked, this toxicity can harm consumers and their freedom to non-discrimination. Tobias is a tool for measuring and exploring toxicity and toxicity bias in language models so that regulators, industry, and the public can minimize discrimination in current and future technology.

Developed state of the art dashboard using PowerBI for several stakeholders in the company. Added data from 3+ data silos into snowflake schema relational database after visiting with stakeholders to determine individual business needs. Created three prototypes that were given to stakeholders to test before providing feedback for next version. Created and deployed new metrics as desired by stakeholders using advanced DAX querying.

Completed end-to-end Data Science project to determine cause of rapid company growth during 2020 during summer internship. Used KNIME visual programming language to clean data that was queried using SQL and DAX from 3+ data silos as well as snowflake schema relational database. The resulting cleaned data was heavily processed before feature selection for Nonlinear Regression model. Model results and discoveries were shown to senior corporate stakeholders in 15-minute presentation. Model was later used by coworkers to write internal white paper on causes of growth and potential methods for maintaining said growth.

As part of the 2022 NFL Big Data Bowl two classmates and I worked on developing a metric focused on special teams plays we named xYAF, short for expected yards after fielding. After cleaning the data using Python we used R to create linear, logistic, and LASSO regression models for xYAF with the goal of quantifying the effectiveness of punt returners against a neutral baseline, as well as examine the circumstances of specific returns.

In order to validate a 2014 paper by Stephen Pettigrew which estimated the effects of the 2013-2014 season NHL realignment three classmates and I recreated the Monte Carlo simulations outlined in the original paper. After using Python to clean and organize the data we used R to simulate the impact the realignment would have on the difficulty of any given team making it to playoffs.