Machine learning applied to quantum computing.
Classical deep learning as the method, quantum computing as the domain, for most of what's here — none of it runs on a quantum computer, see About for why that framing is deliberate rather than a compromise. Two exceptions below, on purpose: causal inference on real ad-exposure data, and an MLOps pipeline on real streaming market data.
WritingA four-part series. Does a learned decoder beat the standard classical one (minimum-weight perfect matching), and under what conditions. It matches given a correct noise model and beats a realistic wrong one by 2x. A fourth decoder, a smaller-scale attempt at DeepMind's AlphaQubit architecture, wins once on real hardware and loses everywhere else.
A GCN trained on free synthetic data, benchmarked against QOBLIB, a suite built by IBM Quantum, ZIB and Purdue to test quantum optimization methods. Wins on 23 of 50 instances, loses badly on one family, and both the win and the loss are reported with the same weight.
Heterogeneous treatment effect estimation on 14 million rows of real randomized ad-exposure data. Meta-learners never beat a naive baseline; a causal forest does, by a real if modest margin. The one project here that overlaps with my actual day job.
A production-style drift-monitoring and auto-retrain pipeline on real streaming market data. Static wins outright, not because retraining is pointless, but because of what happens after the drift it reacts to passes.
Building the data layer that training, retrieval, and evaluation actually depend on. Self-published, 2026. Kindle · Paperback · Hardcover
A Magic: The Gathering side project, separate from the quantum-computing focus above. A card-aware model beats an archetype-only baseline predicting real MTGO tournament results, by a small margin that held up under two separate checks. 13,036 real decklists, real outcomes.