Principal ML Scientist
I turn research into deployed AI
Ten-plus years bridging research and production, across deep learning, reinforcement learning, and scientific ML. I've built search serving millions of shoppers, a climate surrogate that matched dynamical forecast models at a fraction of their cost, and manipulation policies for embodied AI in simulation. Mathematics and physics underneath; production systems on top.
What I Do
I work across the full stack — from research and experiment design to production deployment and operations.
Reinforcement Learning & Embodied AI
RL policies for physical systems: dexterous manipulation, sim-to-real transfer, policy robustness. Isaac Lab, Gazebo, ROS.
Scientific ML
Neural surrogates for expensive simulations, and forecasting on physical and geophysical time series under sparse observation.
Causal Inference & Experimentation
A/B testing infrastructure and causal-inference systems that measure what an algorithmic intervention was actually worth.
Agentic AI Systems
LLM-powered agents that plan, reason, use tools, and execute multi-step workflows — with evaluation built in, not bolted on.
Search & Retrieval
Production search and ranking serving millions of users, plus RAG pipelines that ground LLMs in your domain data.
Computer Vision
Custom deep learning models for detection, classification, segmentation, and visual understanding.
Frequently Asked Questions
What kind of ML work do you take on?
Problems where the modeling is the hard part: reinforcement learning and policies for physical systems, surrogate models for expensive simulations, causal inference and experiment design, and agentic LLM systems that have to be evaluated rather than demoed. I work from problem framing through to a system running in production.
What is a surrogate model and when is it worth building?
A surrogate model is a neural network trained to approximate an expensive physical simulation or numerical solver. It's worth building when you need to run the simulation many times — for forecasting, optimization, or control — and the solver is too slow. Well-built surrogates can be orders of magnitude faster than what they replace while staying within a few percent of it.
What does sim-to-real transfer mean in robot learning?
Sim-to-real transfer is the practice of training a control policy in simulation, where data is cheap and failures are free, then getting it to work on real hardware. The gap is closed with domain randomization, careful reward and observation design, and robustness testing — the policy has to survive physics the simulator never modeled.
Why does causal inference matter for ML teams?
Predictive accuracy tells you what will happen; causal inference tells you what happens if you intervene. Product and business decisions are interventions, so measuring their effect needs experimental design and causal methods, not just model metrics. Without that framework, teams ship models and can't say what they were worth.
How do I book a consultation?
You can book a free call directly on this website. Choose a time that works for you and I'll reach out with a video call link.
Have a problem that needs ML?
Let's talk. 30 minutes, no pitch — just an honest conversation about whether machine learning can actually help.
Book a Free Call