Does the data say what we think?
Silent failures, missing outputs, and label validation — the failure modes that survive every standard sanity check and quietly bias everything downstream.
Applied ML · Model validation · Uncertainty quantification
Hi, I’m Triasha — I build machine learning models for engineering systems, and I test whether they can be trusted.
I hold an MS in Aerospace Engineering from Georgia Tech, where I worked in the Aerospace Systems Design Laboratory on surrogate modeling, uncertainty quantification, and the reliability of large-scale simulation pipelines. Before that I spent eighteen months on airworthiness certification for the Rolls-Royce Trent XWB-84 EP, where a model being wrong was not an academic problem.
Most of my work comes back to the same question: when a model produces a number, what would have to be true for that number to mean anything? That question has taken me through solid rocket motor simulations, electromagnetic solvers, aviation emissions pathways, road networks, and equity markets. I also serve as graduate chair of the Women of Aeronautics and Astronautics (WoAA) chapter at Georgia Tech.
Machine learningModel validationUncertainty quantificationSurrogate modelingTime-seriesPython
data
models
validation
decisions
leakage controlwalk-forward validationGaussian processessilent failure detectionuncertainty quantificationsurrogate modelingoperator learningtime-series
leakage controlwalk-forward validationGaussian processessilent failure detectionuncertainty quantificationsurrogate modelingoperator learningtime-series
How I work
Models are easy to build and hard to believe.
The interesting part of applied machine learning is rarely the model. It is the layer underneath: whether the data is what you think it is, whether the validation protocol actually holds anything out, and whether an uncertainty estimate means what the person reading it assumes it means.
I have found a 15,120-case simulation study where 45% of runs failed silently and nobody had noticed. I have published a trading backtest whose headline finding is that the signal does not work. Both are the same instinct: check the thing before you build on it.
Throughline
Four questions I keep asking
Is the model learning or memorising?
Leakage controls, held-out protocols matched to the structure of the data, and tests designed to fail when the pipeline is wrong.
How wrong might this be?
Gaussian processes, surrogate error characterisation, uncertainty propagation, and knowing where an approximation breaks down.
Does it survive contact with reality?
Transaction costs, regulatory thresholds, certification standards — the point where a model either holds up or does not.
Featured work
Selected projects
01 · machine learning · data quality
Diagnosing silent failures in a 15,120-case simulation study
45% of a solid rocket motor design sweep produced no output and no error message — failures invisible to every standard validity check. A naive split scored 99.8% by memorising duplicated geometries; collapsing to 1,680 unique combinations and testing only on unseen geometry gave the honest figure, 96.8% accuracy at AUC 0.986. Source-level diagnosis traced the silence to an uncapped convergence loop.
classificationleakage controldata qualityroot-cause analysis
Read case study →
02 · time-series ML · validation methodology
An honest equity backtest
A monthly cross-sectional equity strategy built to a strict walk-forward protocol and published with its own negative result. Net Sharpe 0.25 (t = 0.81), rising only to 1.07 with costs set to zero. Every specification tested is logged and committed — including the one that looked best.
walk-forwardmultiple testingpandasscikit-learn
Read case study →
03 · surrogate modeling · scientific computing
Surrogate modeling, Gaussian processes & operator learning
Gaussian process, response-surface and radial-basis-function surrogates compared across benchmark functions, with multifidelity datasets, POD feasibility and Operator Inference — focused on where each approximation breaks rather than which one wins.
View projects →
04 · scientific computing · pipeline engineering
A training run that looked converged and was frozen
A mesh-free boundary element electromagnetic solver written from the integral formulation through to code, and the operator-learning surrogate trained on its output. That surrogate ran for 11 GPU-hours with a loss bit-identical to five significant figures — including across a warm restart that raised the learning rate 200×. NaN guards in the forward pass had rewritten an overflow into a finite loss carrying zero gradient. Confirmed from the checkpoint rather than inferred: AdamW held state for 45 of 61 parameter tensors.
PyTorchautogradnumerical methodsverification
Read case study →
05 · modeling under uncertainty · in production use
GREEN TEA: lifecycle & techno-economic model
Well-to-wake lifecycle and techno-economic model of a forest-residue aviation fuel pathway under parametric uncertainty, benchmarked against GREET, CORSIA and 40BSAF-GREET. The resulting model is in active use by the sponsor, the U.S. Endowment for Forestry.
Read case study →
06 · graph modeling · geospatial
Atlanta mobility resilience digital twin
Road-network model of Atlanta built from OpenStreetMap: disruption scenarios, origin–destination travel-time comparison, and accessibility outcomes under stress. Built with OSMnx, NetworkX and pandas.
View project →Methods
What I bring
I like methods that make models useful without pretending uncertainty, assumptions, or scale do not matter.
Machine learning
Supervised learning on tabular and time-series data, CNNs, clustering, anomaly detection, and feature engineering with scikit-learn and PyTorch.
Model validation
Leakage control, held-out protocols matched to data structure, walk-forward evaluation, shuffled-label tests, and multiple-testing discipline.
Uncertainty quantification
Gaussian process regression, surrogate error characterisation, parametric and Monte Carlo uncertainty propagation, and calibration.
Surrogate & reduced-order modeling
Multifidelity datasets, proper orthogonal decomposition, Operator Inference, response surfaces, and design of experiments.
Statistics
Regression and inference, variable selection under large-sample significance inflation, assumption diagnostics, and subsampling consensus methods.
Engineering practice
Python, pandas, NumPy, SciPy, Git, pytest, reproducible pipelines, and working inside codebases other people have to run.
Focus
Applied ML · Model validation · Uncertainty quantification · Surrogate modeling · Time-series · Scientific computing
Six threads, one question: whether a model’s output can be believed, and what would have to be true for it to be.
Beyond research
Patterns also show up outside the lab.
Making playlists, dancing, and painting keep me close to rhythm, texture, and storytelling — the same instincts I bring to systems work.
Playlists
I build moods through sequencing — small transitions, memory, texture, and rhythm.
Dancing
Movement keeps me attentive to timing, flow, and the way systems feel when they change.
Painting
Painting helps me slow down and notice composition, contrast, and visual structure.
Playlist moodboard
A small creative corner for what I am listening to.
Music is part of how I remember phases of work: writing sprints, debugging sessions, long walks, and the feeling of starting something new.
Find me