I build products on top of machine learning. Most of my work lives in the unglamorous middle: taking models a research team has proven in a notebook and turning them into systems that run every day, make money, and hold up under real traffic.
For the last several years that has meant advertising systems — bidding, measurement, and generative content — at eBay and Twitter, working side by side with applied research and ML engineering teams.
01Recent work
eBay, Promoted Offsite. Product lead for eBay's external advertising product, which syndicates seller inventory across Google, Bing, and social channels. I own the ML-driven bidding, targeting, and seller-protection systems that decide how seller budgets get deployed off-platform, and I partner with applied research to turn their models into shipped product. Work I'm proudest of: conceiving and shipping a generative AI listing-quality program (LLM-based title rewriting and automated image cleanup) that lifted performance across tens of millions of listings; redesigning the target-ROAS bidding architecture, including the prediction models, eligibility gates, and controller logic that adapt bids to seller performance; and owning eBay's response to a major change in Google's bidding policy, recalibrating targets across tens of thousands of campaigns.
Twitter, Ads. Owned product strategy for Twitter's multi-billion-dollar ads portfolio across bidding, ranking, and measurement, as the primary product partner to ML research. Co-developed real-time bid optimization models with research; shipped Twitter's first ML-powered campaign planner for enterprise advertisers; rebuilt ads measurement around causal lift rather than basic conversion tracking; and led the migration of measurement systems to privacy-safe architectures under GDPR and the DSA.
02How I work
Close to the data. I write the SQL and Python myself. The questions that matter tend to show up while you're looking at the rows, and they rarely survive the round trip of a request ticket.
Production over notebooks. A model that works in a notebook is a hypothesis. The product is everything after that: the gates, the fallbacks, the monitoring, and the controller logic that decides what happens on the bad days. That's the part I own.
Startup reps, platform scale. I was the eleventh employee at one startup and the fifth at the next, then spent a decade inside platforms moving billions of dollars. The startups taught me to ship with nothing; the platforms taught me what breaks when you do.
03Earlier
LiveRamp. Product manager on the enterprise data marketplace and API platform, joining through the acquisition of Arbor.
Arbor. Fifth employee at an identity-resolution and publisher-attribution startup, acquired by LiveRamp for $140M.
AeroFS. Eleventh employee at an enterprise file-sync startup in Palo Alto, later acquired by Redbooth. First product job.
04Writing
- The Movie "Home Alone" Perfectly Explains Bayes' Theorem latest
- Ship or Science? When to Turn AI Research into a Product
- The Cleveland Browns Keep Losing. So I Built Them an AI to Win More Games.
- Taming the Black Box: How to Ship AI People Can Trust
More on Medium.
05Elsewhere
Technical depth. I write Python and SQL directly so I can evaluate models, run the analysis myself, and have a real conversation with research scientists about what the numbers mean. B.S. Computer Science and B.A. Political Science, The Ohio State University.
Investing. Co-founder of 6CC, an early-stage fund investing in B2B startups.
Before any of this, I spent a year in college working for a member of Congress. It was a good education in how slowly things move when nobody owns the outcome, and it sent me the other direction: to an eleven-person startup in Palo Alto, where everything was somebody's problem and usually mine. I've been building things since. I live in Brooklyn.