I build machine learning systems for markets, graphs, and simulation.

Sole-authored Train Often, Deploy Selectively, a paper under submission to ACM ICAIF 2026 that separates retraining from deployment. Shadow Before Swap warm-refits a challenger off the serving path, advances it alongside a still-adapting incumbent on the same market stream, and promotes it only after delayed labels mature and a paired NLL advantage clears a fixed deadband. Across 48 UTC weeks of paired replay on Binance perpetual futures — two nonoverlapping episodes, three seeds, eight underlyings — it improves NLL by 0.1472% over calendar replacement and 0.0428% over continuous maintenance while cutting deployed-state turnover by 78.4%.

First-authored an accepted ICML 2026 DEMO workshop paper arguing that effective sample size is not a universal uncertainty diagnostic for off-policy evaluation, and that CI width and empirical coverage are stronger cross-estimator signals. Built reward-noise and distribution-shift stress tests that isolate where effective sample size stays stable even as estimator variance and absolute error climb.

First-authored Stigmergic Graph Memory (SGM), a paper under review at AAAI-27 that reframes endpoint selection in many-to-many warehouse pickup and delivery as a congestion-control decision, not just a routing one. SGM is a bounded, decaying graph-memory layer that reads recent execution signals on warehouse nodes and directed edges to rank which source–destination pairs enter the planner, while a feasibility proof keeps every plan collision-free. Across paired benchmarks on five layouts, three fleet sizes, and 25 seeds, it beats two reconstructed many-to-many baselines in all 15 conditions, with throughput gains of 20.5–36.7%.

Built a Swift-to-AWS Amplify sync layer for continuous mobile-health data collection, raising background-collection uptime by over 150% while cutting daily battery use by roughly 65%. Wrote Python and TensorFlow pipelines that process 200,000+ daily multimodal data points into participant-level time-series features for downstream clinical AI research.