A sample of the research I've worked on. The domains vary — reinforcement learning, planning under uncertainty, sensor systems, blockchain, markets — but the emphasis is consistent: methods that hold up, and systems that make it into use.
Robust policies, built with a k-of-n counterfactual regret minimization subroutine, that exhibit non-obvious cautious behaviour in unfamiliar situations without any task-specific safety design.
Explicit-shaping and bandit-based shaping methods that let an agent learn from arbitrary human advice — accelerating when the advice is good, without being distracted from the original task when it isn't.
Efficient planning for sequential decision-making when agents must act on incomplete information — principled belief-state representations, prediction-reward objectives, and performance guarantees.
My PhD thesis: multi-camera tracking systems that allocate their own resources intelligently to hold tracking performance while reasoning about a dynamic scene and hard compute constraints.
A privacy-preserving, auditable distributed ledger enabling confidential transactions and tokens with instant settlement — and no trusted setup.
Graph-based machine learning for detecting anomalous behaviour in large transaction networks, under class imbalance, limited labels, and operational deployment constraints.
Measuring the quality of learned embeddings with tools from topological data analysis — separability, structure, and failure modes that downstream accuracy alone doesn't reveal.
Shorter, more exploratory work — short-horizon market dynamics, code-repository representation learning, and the occasional interdisciplinary detour.
