About Me

I am a machine learning systems engineer and PhD candidate researching how to make underground mining operations intelligent. My work sits at the intersection of ML research and heavy industry — specifically, predicting energy and fuel consumption for mining vehicles like the Sandvik LH518B (battery-electric) and Caterpillar R2900G (diesel).

I collaborate with CanmetMINING (Natural Resources Canada) on diesel fuel prediction research and run experiments comparing physics-based models, classical ML (XGBoost), deep learning (LSTMs, TCNs, Transformers), foundation models (Chronos, TimesFM, Moirai), and physics-informed neural networks — all on real operational telemetry from underground mines.

Beyond research, I am building high-performance systems in Rust, developing AI-powered tools, and working toward deploying ML models that can run on edge hardware inside mining vehicles.

The mining industry has a pattern: when it adopts AI, it does so in the most brittle way possible. Hand-crafted features. Rigid rule-based systems. Physics models with dozens of calibrated parameters that break the moment operating conditions change. Expert systems that encode one engineer's intuition and can never generalize beyond it.

Richard Sutton's Bitter Lesson states that the biggest lesson from 70 years of AI research is that general methods leveraging computation are ultimately the most effective — by a large margin. Methods that try to build in human knowledge work in the short term, but plateau. Methods that scale with data and compute keep getting better.

Mining has not learned this lesson yet. Most mining AI projects fail not because ML does not work, but because the implementations are designed to be brittle. They bake in assumptions about specific sites, specific vehicles, specific operating conditions — and then break when anything changes.

My thesis tests this directly: I compare every modeling paradigm — from first-principles physics to self-supervised foundation models — on the same mining data, and measure which approaches actually generalize. The question is not “does ML beat physics?” It is “under what conditions does each paradigm excel, and what scales?”

I am building toward becoming an Industrial ML Specialist — the person who can take raw telemetry from any heavy vehicle fleet, pre-train self-supervised representations on unlabeled data, fine-tune foundation models on the small labeled portion, inject physics constraints where they help, implement the whole thing in high-performance Rust, and deploy it as a monitored production system on edge hardware.

That means going deep on three fronts simultaneously: advanced ML research (self-supervised learning, physics-informed neural networks, world models), high-performance systems engineering (Rust, CUDA, edge deployment), and production AI infrastructure (agents, MLOps, local LLM distillation).

The person who combines all three of these capabilities in the context of mining does not exist in the industry today. I intend to be that person.