I bring applied-math rigor — numerical analysis, dynamical systems, probability — to modern AI: how sequence models actually work, how to evaluate them honestly, and how to keep the claims auditable. These threads run through the work, with causal inference and evaluation as the methodology spine.
Why do state-space models (Mamba, S4, and kin) behave the way they do? I trace it back to the numerics — discretization, stability, dynamical systems — and write it up as a lens-led foundations book, in progress.
The methodology spine: double machine learning for time series, leakage-safe temporal cross-validation, and causal implementations in Julia and Python built toward reference quality.
Do detectors hold up out of distribution? Honest evaluation — baseline ladders, bootstrap confidence intervals, calibration, and stop-gates. These are methodology proofs-of-concept, not production security claims.
Bridging classical optimal control (Bellman, dynamic programming, LQR/MPC) and modern reinforcement learning (policy gradient, actor-critic, model-based) — a structured guide, in progress, with a live citation graph over the literature.
Every substantive claim is meant to be checkable. A local knowledge base
(research-kb) and research workflows (research_toolkit, in
development) back the citations, and substantive projects carry a decision log and an
independent audit where the work warrants it.