Experience

At PureSkill.gg, I co-founded and scaled an automated coaching platform for competitive gamers. The goal was to detect player mistakes from gameplay data and direct players toward focused practice that would help them improve. This was one of the most technically creative roles I’ve had because it involved turning noisy, high-volume game telemetry into useful coaching feedback. I built LightGBM models that estimated win probability hundreds of thousands of times per match so we could measure the impact of player decisions in a quantitative way. I also built models to evaluate things like smoke grenade quality and counter-strafing technique using clustering, regression, and other machine learning methods. Just as important, I designed the system that prioritized coaching recommendations based on win impact, frequency, and difficulty, while adjusting for player rank so the advice stayed relevant to the player’s skill level. We processed over a million matches and built a product that served thousands of players, including paying customers. I also worked with human coaches to calibrate the system so the automated feedback matched real player expectations and coaching standards.

Alongside the core product work at PureSkill.gg, I also contributed to the broader gaming analytics community. I released public data on AWS and supported it with documentation, tutorials, and open-source software so other researchers could work with it more easily. I also provided modeling advice that contributed to published research on how feature importance in win-rate models differs across skill levels, from professionals to newer players. In another project, I advised a master’s student on how to convert complicated player telemetry into a more understandable vector space using methods like Word2Vec. I liked this part of the work because it combined technical problem solving with teaching, collaboration, and helping other people do better research.

In my current role on the Pricing team at Kohl’s, I lead work on large-scale forecasting and pricing systems. One of the biggest things I’ve owned has been the delivery of a deep learning pricing system built around Temporal Fusion Transformers, which helped support better discounting decisions and drove significant revenue impact. A big part of this role has been bringing structure to a growing team by setting coding standards, creating review processes, and improving how we track experiments so models can move into production more safely and consistently. I also serve as a key business-facing partner for pricing model launches, which means I’m not just building models, but also helping decide when they are ready to go live and how they should be used. On top of that, I’ve mentored junior teammates and helped improve production model accuracy through better feature work and model development practices.

Before joining Pricing, I worked on the Search and Digital Signs teams at Kohl’s. On the search side, I owned personalization modeling and built logistic regression models that drove measurable revenue gains through A/B testing. I also helped move the team toward embedding-based retrieval by building product and query vectors that enabled semantic search across a very large catalog. Beyond modeling, I spent time improving pipeline efficiency by refactoring older Scala workflows into Spark, which made the codebase much smaller and the jobs faster to run. I also built an LLM-powered classifier for customer feedback that reduced incident response time by surfacing important issues much faster. On the digital signs side, I designed and rolled out LLM-generated marketing content at very large scale across hundreds of thousands of in-store signs. That work required more than just prompt writing. I built guardrails, evaluation frameworks, and validation processes, and I partnered closely with Legal and AI Governance to make sure the system could be deployed safely and responsibly.

At Rho Impact, I worked on NLP and recommendation problems. One of my main projects was building a TF-IDF and logistic regression system that classified a very large volume of medical documents with strong accuracy. I also developed an investor-company recommender based on collaborative filtering. What I liked about this role was the variety. It gave me the chance to work on both text classification and recommendation systems in a practical setting where the models had to be useful, not just technically interesting.

At Lowe's, I worked on retail optimization and data engineering problems. I built a greedy assortment optimization algorithm designed to increase sales while reducing waste, which meant balancing business value with operational constraints. I also built a data pipeline using Hadoop and SQL to estimate stock levels from messy retail data. That experience gave me a strong appreciation for the importance of solid data foundations. A lot of the value in data science comes from modeling, but a lot of it also comes from being able to make sense of incomplete, inconsistent, or hard-to-use data and turn it into something reliable enough to build on.

My research background is in physics and astronomy, where I spent several years studying galactic outflows at high redshift. During that time, I published a first-author paper, designed open-source Markov Chain Monte Carlo software for decomposing galactic emission signals, and helped process data from one of the largest surveys ever conducted with the Keck Telescope. I also spent more than 15 nights taking telescope data as part of a careful observing plan and developed software to extract usable information from raw astronomical images. Just as importantly, I trained other team members to use those tools. That period of my career shaped a lot of how I work today. It taught me how to deal with noisy data, how to build reliable analytical tools, how to communicate technical ideas clearly, and how to stay patient with hard problems that take real persistence to solve.