Classical deep learning on quantum computing problems, and causal inference on real ad data
Six years building and shipping production ML and agentic systems inside enterprise environments at Amazon. PhD research on formal verification and evaluation of AI systems. Most of the projects here use classical deep learning on problems drawn from quantum computing: the domain is quantum, the method is not, and none of it runs on a quantum computer — see the write-ups for why that's deliberate. Two exceptions, on purpose: a causal-inference project on real ad-exposure data, the actual discipline behind my day job, and an MLOps drift-monitoring pipeline on real streaming market data.
| qec-neural-decoder | When a learned decoder beats minimum-weight perfect matching, and when it doesn't. Four architectures now, including a smaller-scale attempt at DeepMind's AlphaQubit — wins once on real hardware, loses everywhere else. Models and dataset on Hugging Face. |
| neural-mis | A GCN vs. classical greedy on QOBLIB's Maximum Independent Set benchmark, wins concentrated exactly where greedy was weakest, one family it loses badly and stays that way after two documented fix attempts. |
| uplift-modeling | Heterogeneous treatment effect estimation on 14M rows of real randomized ad data. Meta-learners lose to a naive baseline; a causal forest doesn't. Model and a synthetic ground-truth benchmark on Hugging Face. |
| drift-monitor | Production-style drift detection and auto-retrain pipeline on real streaming market data. A model that's never retrained beats both a periodic and a drift-triggered retraining strategy — the mechanism traces to regime reversion, not to retraining being pointless. |
| toolbudget | Cost-aware benchmark for tool-using agents: what the answer cost, not just whether it was right. |
mtg-deck-matchup, a Magic: The Gathering side project separate from the quantum-computing focus above. A card-aware model predicts a real MTGO Standard decklist's tournament performance better than its archetype label alone, on 13,036 real decklists joined to real outcomes. A small effect that held up under two separate checks. Dataset on Hugging Face.
Data Infrastructure for AI Systems — building the data layer that training, retrieval, and evaluation actually depend on. Self-published, 2026.
