Est. 2007 · Berkeley, California
Signal extraction where the noise dominates.
Statistical Machine Learning in Low Signal-to-Noise Regimes. We develop state-space estimators, shrinkage methods and reproducible research infrastructure for non-stationary, high-dimensional data.
Research programme
Three problems, one discipline.
Non-stationary state estimation
State-space models and adaptive filters for latent quantities that drift, switch regimes, and are observed only through heavy measurement noise.
High-dimensional estimation
Shrinkage, regularization and random-matrix theory for covariance and regression problems where the number of variables is comparable to the sample size.
Compute-efficient inference
Deterministic, auditable research pipelines: seeded computation, content-addressed artifacts and benchmarks that can be reproduced by anyone.
Methodology
Optimal weighting, made explicit.
In a scalar random-walk model with process variance and observation variance , the steady-state Kalman gain depends only on the signal-to-noise ratio. When , the gain scales with the square root of the ratio, which fixes the effective averaging window of the estimator.
Steady-state gain of the scalar Kalman filter for a random-walk latent level.
Research & engineering blog
Recent publications
Methodology
Kalman Filtering for Latent Signals in Low Signal-to-Noise Regimes
A derivation-first walkthrough of the Kalman filter as the conditional-mean estimator for a latent state observed through heavy measurement noise, with an explicit account of how process variance controls tracking of non-stationary dynamics.
October 1, 2026 · 3 min read
Methodology
Shrinkage for High-Dimensional Covariance Estimation
Why the sample covariance matrix becomes unreliable when the number of variables is comparable to the number of observations, and how linear shrinkage restores conditioning with a data-driven intensity.
September 16, 2026 · 2 min read
Engineering
Deterministic Research Pipelines: Seeds, Fingerprints and Typed Configuration
A practical design for research code in which every result can be traced to an exact configuration, a seed and a content-addressed artifact, so that any estimate can be re-derived by a colleague months later.
August 27, 2026 · 3 min read
Publications on this site are educational and research material. They are not investment advice and do not solicit investment from the public. Read the terms of use.