Two systems I built — one in systematic trading infrastructure, one in multi-model risk prediction.
A modular Python framework for quantitative portfolio management: the full stack from factor estimation to constrained optimization to walk-forward execution, with every step grounded in theory and covered by automated tests.
The constrained MVO solves the standard mean-variance problem with active weight penalties and a full constraint set:
For spiked covariance structures (low-rank factor model + diagonal idiosyncratic), the platform uses the Woodbury-Sherman-Morrison identity for O(n) solves instead of inverting the full \(n \times n\) matrix:
A hierarchical ensemble in which LSTM, graph neural network, and probabilistic models each read one physiological subsystem, and belief propagation fuses their outputs into a single risk assessment. The application is clinical (early detection of preeclampsia), but the architecture is domain-agnostic: it's the same fusion pattern that applies to any decision built from several partial, noisy signals.
Across four subsystems it reached an AUC of 0.963 on 2,847 patients at three sites, against a 0.890 baseline. In a prospective trial of 500 patients it caught events 14.3 days earlier than standard care (p < 0.001). INT8 distillation then shrank the model by 93% while keeping accuracy within 1.2% of the original.