work Cutting search cost in computer-aided synthesis planning A learned cost-to-go heuristic for retrosynthetic search. Reordering candidate reactions by their remaining distance to purchasable materials raises the solve rate at unchanged per-step compute. Industrial R&D project, owned end to end. Bayesian optimization that has to pay for its experiments A closed-loop optimization system for self-driving laboratories where every proposed experiment is priced, charged to a ledger, and refused if the budget cannot cover it. Industrial R&D project, owned end to end. A reproducible handover between two machine learning subsystems An installable, versioned component that decides which molecular descriptors describe a property, and whose output is required by test to reproduce the validated result exactly. Industrial R&D project, owned end to end. research Regularized Nyström curvature for large-scale optimization Newton's method costs O(nd² + d³) per step. Sampling a few columns of the Hessian and reconstructing it with the Nyström method brings that to O(ndm + dm²), and a regularizer keeps the step usable when the Hessian is indefinite. Variance reduction with Barzilai-Borwein step sizes SVRG removes most of the noise from stochastic gradients but still needs a step size. The Barzilai-Borwein rule derives one from gradients already computed, making it essentially free curvature information. Second-order federated learning on edge devices Newton-type methods converge in far fewer communication rounds than FedAvg, and nobody uses them because a Hessian for a 25M-parameter model does not fit on a client. A sampled low-rank sketch changes that. learning Credit risk classification A self-directed exercise on public UCI credit datasets. End-to-end supervised pipeline with class imbalance handling, cost-sensitive evaluation and model interpretability.
