I develop optimization and data-driven methods for fair and reliable prediction, infrastructure modeling and planning, and scalable computation.
Property assessment, the resource footprint of data centers, and first-order solvers on GPUs look unrelated from outside. Each is a computation that a decision rests on, and in each the useful work is being precise about what that computation can and cannot support.
Predictive systems whose errors carry distributional or institutional consequences — where accuracy alone is not the right objective.
Energy, water, and environmental systems reconstructed from imperfect public records, with an explicit boundary between reported and modeled quantities.
Decomposition, sparsity, and first-order methods on modern hardware, with numerical reliability that survives scale.
The three directions I am working on now, each with a public repository.
How can mass-appraisal models improve predictive performance without worsening systematic vertical inequity across properties?
A mass-appraisal model can be made more accurate and still leave some owners systematically over-assessed; the two goals have to be optimized together.
How can public data support transparent, reproducible modeling of the infrastructure and resource externalities associated with large data centers when detailed operational telemetry is unavailable?
What a large data center draws from the grid, the water system and the aquifer around it can be partially reconstructed from public records — provided the reconstruction is explicit about which quantities are reported and which are modeled.
How can first-order optimization methods remain reliable when large-scale solvers are implemented on GPU hardware?
A first-order solver that runs fast on a GPU is only useful if its termination criteria still mean what they meant on a CPU.
Cook County Assessor's Office
Merged support for the covariance-penalized LightGBM objective developed with the MIT research team into its open-source residential assessment model.
The objective is selectable in the model configuration as mse_cov. This
records a change to a public codebase, not a measured effect on assessments or
policy.
Research repository
Research code
Research software
Conference presentation (unpublished)
PhD student at the MIT Operations Research Center, since September 2024.
I came to MIT from the Universidad de Chile, where I completed a master's in operations management and a bachelor's in industrial engineering. My master's thesis — a column-generation decomposition for large-scale feature selection — is where the interest in scalable convex optimization started. Between degrees I worked as a research engineer at Nezasa AG in Zurich on time-dependent routing, and with the Web Intelligence Centre and ACHS on demand forecasting and capacity planning for a hospital in Santiago.
I am advised by Saurabh Amin and Deep Deka. Earlier in the PhD I also received research guidance from Haihao Lu. Before MIT I worked with Fernando Ordóñez and Renaud Chicoisne on decomposition methods for feature selection, and with Charles Thraves and Ricardo Montoya at the Complex Engineering Systems Institute.
Outside research: music, climbing and time outdoors.
Published a public-data modeling repository for data-center infrastructure and resource systems, beginning with a transparent Prineville case study.
The Cook County Assessor’s Office merged support for the covariance-penalized LightGBM objective developed with our MIT property-assessment research into its open-source residential assessment model, as pull request #475.
My master’s thesis received a thesis distinction from the Universidad de Chile’s School of Graduate Studies.