About

I learned the methods through mathematics and econometrics, then learned the work through client decisions.

What I do

I'm a marketing data scientist and consultant with nearly five years of experience. Most of my work sits where marketing, customer behavior, and data systems meet: marketing mix models, forecasting, experiments, NLP, document retrieval, and the data plumbing that makes those methods usable week after week. I like both halves of the job, talking through model internals with technical people and telling a business owner what a result means for them.

I studied at the University of South Florida and earned a B.A. in Mathematics and a B.S. in Econometrics and Quantitative Economics. I left knowing R, Stata, and a lot of statistics, and taught myself Python and SQL because that's what industry runs on. A C++ scientific computing library with Python bindings, SicLib, helped get me my first data science job.

Ringer Sciences, 2021 to 2024

From November 2021 to February 2024 I was a Data Scientist at Ringer Sciences, a marketing analytics consultancy with a small, hands-on team. I worked across six client engagements, and it's where I learned marketing data science and client management.

I built and ran sentiment analysis, named-entity recognition, and text classification on more than 2 million social posts in daily pipelines. Client teams changed media and content plans based on the findings, and the work became a product sold to other clients.
For a large hospital client, I pulled hand-typed journal ad data from Google Sheets into BigQuery, joined it with ad platform data, and built dashboards for marketers and physicians. They became a recurring paid product. It's still my favorite project because it ran end to end, from messy spreadsheets to takeaways doctors could use.
I built a retrieval assistant over a B2B media client's own reports. Analysts spent about 40% less time on research, and the client kept it as a recurring paid product.
I developed marketing mix models, forecasting models, and A/B test frameworks that teams used to set channel budgets.
I automated 12 recurring analyses, saving analysts about six hours a week, and packaged deliverables so repeat work needed up to 90% less manual effort.

Independent consulting, 2024 to now

Since February 2024 I've worked with food and beverage, agriculture, and consumer businesses. In a small business, I'm often the only analytics person, so the job runs from campaign setup to the owner's decision.

I tied paid campaigns to verified visits with dated flyers, phone numbers, and personal offer codes. Cost per attributable revenue dollar on lead campaigns fell about 70% after I rebuilt the audiences from real visitors.
I found demand for a new event format by analyzing Reddit, Meetup, and group chats. Foot traffic doubled within one quarter of the launch.
I built a lead CRM for a wholesale supplier's sales team and a random-forest ranker on top of it, deployed on Google Cloud's Vertex AI. Closing rate per call rose about 50% once the team called the strongest leads first.
For another client, I scored prospects from business records, and lead conversion rose about 30% once the scoring chose who the team contacted.
I raised average ticket about 30% with a two-pastry bundle picked from POS basket analysis.
I train Bayesian marketing mix models, run Google Ads and Meta Ads funnels, and forecast visits by hour so owners know what a given spend means for staff and stock.

AI systems and agents

I use language models where they save writing or search time, and I keep them away from the math. Calculations happen in Python against the source data, and the model writes around numbers that are already computed. My economic analytics project shows the pattern: after I moved the calculation out of the model, all 21 current insights passed a database check.

I also build and direct agent systems:

Agent telemetry. I built a system that tracks cost and error rates across single-agent and multi-agent workflows and recommends cheaper setups. For one business, token use per task fell about 70%, with lower error rates.
A recommendation engine. I built the recommendation engine for a multi-source web data ingestion platform. Records pass through rule-based filters, then an ML model, then an LLM reviewer. The first model distilled the LLM reviewer's labels into a LightGBM model over sentence embeddings, chosen from three model families against a rule baseline with grouped cross-validation, then validated once on a later held-out window: AUC 0.74 against 0.53 for the rule, with a 95% interval of the gain from 0.13 to 0.29. It scored about 7,700 live records in 72 ms each at no marginal cost, against roughly $0.05 and 0.8 seconds per record for the LLM reviewer. A two-stage classifier has since replaced it. I also priced four review policies by LLM spend plus the cost of missed and false top records. In replay, the winner, an MLP filter ahead of the LLM, kept 60 of 74 top records with 719 LLM reviews instead of 1,707. I designed and directed the engine, and AI coding agents wrote most of the code under my review protocol.
A strategy game. For an in-development historical strategy game, I designed the workflow the coding agents build under and run the player and market research.

I use Claude Code and Codex as development assistants. They handle routine ETL and dashboard wiring. I own the analysis, the data transformations, and the design, review what they produce, and decide which findings matter to the business.

Outside work

I paint while traveling because it makes me look at a place more slowly than a camera does. The painting pages include work from Mongolia, Kyrgyzstan, Chile, and Mount St. Helens.