Projects

Embeds first-principles thermal dynamics constraints directly into Bayesian neural network priors, producing calibrated uncertainty estimates for building energy prediction. Improves predictive accuracy 15–27% over purely data-driven baselines in data-scarce regimes. Published in Forecasting, 7(1), 9, 2025.

Proposes a principled framework combining mutual information criteria and Bayesian model evidence for ranking competing surrogate model architectures. Consistently identifies parsimonious models that generalize better than cross-validation alone across 20 real-world building datasets. Published in Information Sciences, 726, 2026.

Comprehensive review of 150+ UQ methods applied to ML-based building energy models. Establishes a taxonomy of uncertainty sources, synthesizes evaluation standards across studies, and proposes a standardized evaluation framework. Published in Renewable and Sustainable Energy Reviews, 218, 2025.

Agentic AI system that orchestrates building diagnostics, energy simulation, and equity analysis agents to generate equitable, cost-effective retrofit recommendations for residential buildings, with explicit attention to energy burden disparities across income groups.

ML-based diagnostics framework for identifying cost-effective, equity-aware retrofit measures in residential buildings, with particular attention to energy burden disparities experienced by low-income households.

Deep learning pipeline for automated extraction of spatial features from architectural floor plan drawings, enabling programmatic building energy model geometry generation directly from 2D architectural drawings.

SM-2 spaced repetition system for LeetCode problems. Surfaces due reviews, tracks mastery per topic, and visualizes learning curves over time — fully client-side with zero dependencies, deployable as a static page.