Exact optimization methods for systems that must commit capacity before demand is known, and algorithms that make those models solvable at operational scale.
Scholarly work
Research
Two-stage stochastic optimization for outpatient clinic scheduling under uncertainty
2023 to present · with T. Giovannelli, K. S. Shehadeh, and J. Shi · Ph.D. dissertation, manuscript in final co-author revision
Stochastic programming
Enhanced integer L-shaped method
Branch-and-cut
Healthcare
Question
How should a clinic book a session when a sizeable share of patients will not attend as scheduled, some cancel at short notice, and others arrive without an appointment?
Approach
A two-stage stochastic program with an integer first stage, solved exactly by an enhanced integer L-shaped decomposition inside a branch-and-cut algorithm.
Outcome
Booking policies a clinic can act on directly, covering how far to overbook, whether to hold slots open for same-day demand, and how to route walk-ins across providers. Results show that once a session is busy, double-booking outperforms reserving open-access slots, and most of the same-day access comes from letting walk-ins move between providers.
Allocating open-access appointment slots in a single-provider session: a two-stage stochastic multi-objective program with an integer first stage
2023 to present · with T. Giovannelli · Ph.D. dissertation, manuscript in preparation
Multi-objective optimization
Set-valued optimization
Aumann integral
Healthcare
Question
Clinics cannot credibly trade off patient waiting time against provider time, yet a single-objective scheduling model fixes that trade-off through predetermined weights before any schedule is produced.
Approach
The nondominated set of a multi-objective recourse problem is treated as a random closed set and minimized in the set-valued sense, over a binary first stage.
Outcome
The full set of efficient open-access allocations, so a manager chooses a schedule by the trade-off they are willing to accept rather than by weights they had to invent.
Virtual metrology for chemical mechanical planarization process in semiconductor manufacturing
2023 to 2024 · with X. Han and X. Jia · Published in the ASME MSEC 2024 proceedings
State-space filtering
Semiconductor manufacturing
Predictive modeling
Question
Material removal rate cannot be measured on every wafer, so process drift goes unseen between physical measurements.
Approach
A dual linear Kalman filter that infers removal rate from in-process signals, developed at the Lab for Intelligent Metrology Systems.
Outcome
Wafer-level estimates accurate enough to act on, outperforming multiple linear regression, EWMA, and Bayesian ARX benchmarks.
Multi-task learning for time-to-event modeling of post-heart-transplant outcomes
2021 to 2023 · with X. Chen, M. Gentili, and J. Trivedi · Presented at INFORMS 2022 and Research!Louisville 2022
Time-to-event analysis
Multi-task learning
Proportional hazard model
Health services
Question
One-year survival after heart transplantation differs across centers, but pooled models cannot separate the center's contribution from the risk profile of its patients.
Approach
A multi-task Cox model treating each center as its own task over a shared representation, combining national registry records with neighborhood deprivation measures around each center.
Outcome
Center-level effects isolated from patient-level risk, identifying where system-level intervention, rather than case selection, would change outcomes.
Predictive modeling from limited, uncertain, and imprecise data: grey-theory-based neural networks for supply chain performance prediction
2018 to 2020 · with S. M. Ali, G. Kabir, and S. K. Paul · Published in Sustainability, 2024
Supply chain
Forecasting
Neural networks
Grey systems theory
Question
Supply chain demand and performance must often be predicted from short, noisy histories where classical statistical forecasting is unreliable.
Approach
Grey-system models coupled with neural networks, developed in the M.S. thesis and extended into a GM(1,1)-based grey neural network architecture.
Outcome
Forecasts with sharply reduced mean relative error across key performance indicators, usable for planning where conventional methods have too little data to fit.
I intend to pursue four research directions, each centered on methodological questions that I investigate within outpatient care delivery, air cargo networks, and manufacturing operations.
01
Networked capacity with shared resources
I want to understand how to allocate capacity across sites that draw on a shared resource pool, such as providers in a clinic network, aircraft and crew in a cargo network, or production lines across plants. Routing demand between sites links the recourse subproblems that decomposition normally separates, so the methodological question is how much service improvement cross-site routing actually delivers relative to the extra solve cost it creates.
02
Exact decomposition at operational scale
I am interested in how far exact methods can be pushed when the instances reflect an entire planning horizon. The research question is which combinations of stronger and aggregated cuts, warm starts from heuristic policies, and parallel scenario evaluation make integer L-shaped and related approaches viable at operational scale. The aim is to identify techniques that generalize to any two-stage program with mixed-integer recourse.
03
Efficient sets for multi-criteria decisions
I am interested in developing methods that replace fixed weights with explicit trade-off surfaces. A weighted objective forces a trade-off before any solution exists, whether the terms are patient waiting versus provider time or service level versus cost. The question I want to investigate is how to compute the efficient set directly, so the operating point is chosen afterward, and criteria like access and equity, which expected-value objectives cannot represent, appear in the output rather than being pushed into weights.
04
From model to decision support
I am interested in how to move from a model to a tool that planners can actually use. The questions include how to calibrate stochastic parameters from operational data, how to backtest recourse-based policies against realized outcomes, and how to embed these models in decision support systems that practitioners run themselves. My applied work in air cargo network planning and consumer goods manufacturing has shaped this agenda.
Applied work
Projects
Summer 2026
Stochastic evaluation framework for air cargo network planning
Atlas Air, Network Planning
Scenario-based evaluation with mixed-integer recourse for planning under operational uncertainty, covering fleet flow, capacity utilization, routing, spill, and cascading delay.
Disruption cost quantified through Monte Carlo evaluation of the recourse value function.
Stochastic parameters calibrated from operational data on Azure Databricks, with automated validation and end-to-end run provenance.
Route network represented on Amazon Neptune for the connectivity queries feeding the model.
Pyomo
HiGHS
Gurobi
Monte Carlo
Azure Databricks
Amazon Neptune
2024 to 2025
Production planning automation for Personal Care manufacturing
P&G Digital Accelerator, University of Cincinnati
Databricks pipelines translating SAP planning volumes into production-line recipe rates.
Roughly 550 hours of annual manual effort removed, about half the prior workload.
Databricks-to-Tableau integration automated through KNIME, with deployed KPI dashboards.
Azure Databricks
Python
SQL
KNIME
Tableau
SAP
2023
Cybersecurity workforce training for the logistics industry
Logistics and Distribution Institute, University of Louisville
Authored five online, self-paced instructional modules: IT Basics, Network Foundations, Network Security, Risk Analysis, and Capstone.