About
An engineer and researcher who follows questions into whatever has to be built to answer them.
The practice
Most of what is on this site began as something I wanted to know and could not look up. Whether wavelet-conditioned models can extrapolate a price series. Where in a market's timeframes learnable signal lives, if it lives anywhere. Whether a contested published tomography method does what its papers say it does. None of it was assigned. It was interesting, so I built what the question needed and wrote down what came back.
The range is not a strategy. Tropical-Pacific sea-surface temperature sits beside order-book microstructure, synthetic-aperture radar, and a century of Toronto school buildings because each was, at some point, the thing I was curious about. What holds across them is a willingness to build whatever the question needs rather than shrink the question to fit an existing tool — network layers and training loops written against an automatic-differentiation API, a nine-estimator robust-statistics library, an exactly invertible wavelet layer, CUDA and OpenMP reconstruction code.
The other half is a temperament about evidence, which the engineering training sharpened rather than invented: an interest in what would have to be true for a result to be wrong, and in saying plainly how much weight it can take. Exploratory does not have to mean loose. Most of the studies here declare their gates before the run, keep something simple in view for comparison, and report the answer at the size it arrived in.
Curiosity picks the question; the discipline is what makes the answer worth having.
Habits
They recur whether or not a project calls for them. Each links to a study where you can watch one at work.
What result would settle the question, written down before the work starts. It is the difference between exploring a question and wandering around one.
A pre-registered scan asked which timeframes and horizons carry learnable signal at all, before any model was committed to.
Signal-Scale ScanEvery representation, label, and model class is measured against something simple and interpretable. A number with nothing beside it is not a result.
Matched baselines beat the neural models at every horizon, and the wavelet representation made them worse rather than better.
WaveGAN-2A result's assumptions, failure modes, and evidence tier stay attached to it wherever it goes, rather than being filed in an appendix nobody opens.
A written charter fixes the metric definitions and the evidence hierarchy for an eleven-document series, and records the estimator boundaries across which numbers must never be ranked.
Series Charter and Metric ProtocolOverclaiming is the expensive error. A study that narrows the space has still established something, and it is worth stating at the strength it actually has.
A quantile head that settled on a near-constant median turned out to be reporting the structure correctly: the predictable quantity was scale, and permutation controls located it precisely.
Frozen-Backbone Head LabThe range
Different subjects, different tools, different literatures — grouped here by where the evidence actually sits.
Research software and empirical studies: forecasting systems, invertible signal representations, GPU-accelerated reconstruction, and the evaluation protocols that decide whether any of it worked.
Evidence Wavelet-domain ENSO forecasting, regime-adaptive probabilistic forecasting, Doppler tomography
Questions about buildings, housing, and public infrastructure answered with archival sources, interviews, site observation, and structured records rather than assertion.
Evidence Architectural Conservancy of Ontario, Ottawa Community Housing
Field records, drawing review, inspection workflows, and the documentation continuity that lets a multi-party project follow its own issues and decisions.
Evidence MMM Group / WSP municipal infrastructure, Ottawa Community Housing
Turning scattered source material — methods, field information, literature — into something a reader can follow, check, and act on. Twenty-plus published reports behind it, and eight years of tutoring before that.
Evidence Published report series, tutoring, methods documentation
The short version