Frame before you fit
The hardest part of ML is rarely the model. I spend real time defining the problem and the metric before training anything.
My story
From physics labs to production ML.
I started in computational physics, where I learned that a beautiful equation is worthless if you can't measure it against reality. That obsession with rigorous evaluation never left — it just moved from particle simulations to machine learning systems serving millions of people.
I've led data science at a fintech and an e-commerce platform, built ML teams from scratch, and shipped models that touch fraud, personalization, forecasting, and vision. In 2021 I started taking on independent engagements so I could stay close to the hard problems — no hand-offs, no diluted intent, just the model and the metric.
When I'm not training models, you'll find me bouldering, tinkering with mechanical keyboards, and losing arguments to my cat about whether the heater counts as a workstation.
Skills & tools
How I work, and what I work with.
Core proficiency
Stack & tooling
Experience
Where I've worked.
-
2021 — Present
Independent Data Scientist & ML Engineer
Signal · Toronto & Remote
Partnering with startups and scale-ups on end-to-end ML — problem framing, modeling, deployment, and monitoring. 60+ projects across risk, personalization, and forecasting.
-
2018 — 2021
Lead Data Scientist
Verde Pay
Built and led a 5-person ML team. Shipped the real-time fraud detection platform that cut losses by 38% and introduced the company's first model monitoring stack.
-
2016 — 2018
Machine Learning Engineer
Atlas Commerce
Owned the recommendation and ranking systems for a marketplace reaching 40M monthly users. Drove experimentation, feature engineering, and serving infrastructure.
-
2014 — 2016
Data Analyst & Research Assistant
University Computational Physics Lab
Built simulation pipelines and statistical models for high-energy physics data, co-authoring two papers and discovering a love for turning noise into signal.
What I believe
Principles that guide the work.
Honest about uncertainty
A calibrated maybe beats a confident wrong. I report error bars, failure modes, and the limits of every model I ship.
Production is the point
A notebook isn't a result. I design for serving, monitoring, and the day the data drifts — because it always does.
Simple baselines first
Logistic regression has shipped more value than most deep nets. I earn complexity, I don't assume it.
Reproducible by default
Versioned data, pinned environments, tracked experiments. If I can't rerun it, I don't trust it.
Communicate the so-what
Stakeholders don't need ROC curves — they need decisions. I translate models into outcomes people can act on.
Let's make the data useful.
If you're building something that needs machine learning to actually work in production, I'd love to hear about it.