I'm a machine learning engineer working on independent applied AI research. I'm interested in how models store information and reason with it, and in how we evaluate and interpret whether they're doing what we intend.
I came to ML from lab research and spent six years building ML products on production systems. At Red Bull Media House I owned video and article recommendation systems end to end. At Tinder I was a lead engineer for paywall personalization on the ML Revenue team, where my work centered on uplift modeling, ranking, and experimentation.
The failures I saw most in production were upstream of the model: reward signals that didn't reflect what we actually wanted, and training data that only captured a portion of our desired behavior. That's where I learned that the key questions are often the simplest to ask and the hardest to answer: are we solving the right problem, and are we measuring the right thing?
Outside of work you can find me at the crag or in the mountains.
Open to research engineering and applied ML roles.
CURRENT WORK
A research proof of concept on how a structured specification, built from retrieved practitioner knowledge, affects agentic execution on ML engineering tasks. The pipeline, pre-registered experimental design, and MLE‑bench evaluation harness are built and calibrated.
A Cohere Labs community project asking whether social stereotypes surface in a model's internal computation before they show up in its output. I built the batch lens read-out and a template lens for multi-token targets, then tested lens swaps against counterfactual patching on BBQ. Findings are preliminary.