I'm Juan Pablo Morande, a PhD candidate in Industrial and Systems Engineering at Auburn University. I work on logistics, scheduling, and forecasting problems using optimization and machine learning.
PhD, Auburn University, Spring 2027
Applied Research Engineer roles in the US. Open for relocation. Available to start in 2027.
I'm a fifth-year PhD candidate in Industrial and Systems Engineering at Auburn University. My research covers operations research, machine learning, optimization, and simulation, with a dissertation on integrating prediction and optimization for complex logistics problems. I've been on the INFORMS student chapter board every year since 2022, including a term as President in 2024–25 and currently serving as the E-council representative.
My projects have included two-dimensional truck loading optimization, military aircraft and crew scheduling for U.S. Air Force wargaming simulations, deep learning applied to illicit supply chain detection, freight rate forecasting with explainable AI, and drone-based order-picking in warehouses. In 2024, I spent a semester as a visiting researcher at NTNU in Norway, developing closed-loop supply chain models. More recently, I've worked on LLM-based systems, including a multi-agent framework for automated algorithm design and a fine-tuned RAG system for legal advising. This summer I'm an Applied Scientist intern at Amazon, working on routing for sub-same-day delivery. I'm designing enhancements that balance short-term planning with real-time dispatching, with the goal of rolling them out across North American and European delivery stations.
I'm originally from Chile and fluent in both Spanish and English. I got into research through humanitarian logistics simulation at Universidad de los Andes, and I'm now looking for an Applied Research Engineer role where I can put optimization, data science, and ML to work on real problems.
Forecasted shipping freight rates in Chilean ports using Temporal Fusion Transformers, LSTMs, and attention-based architectures, with LIME and SHAP for interpretability and Bayesian methods for hyperparameter tuning. The explainability analysis surfaced the main drivers behind rate changes for maritime logistics stakeholders.
Developed optimization models and heuristics for U.S. Air Force wargaming simulations, scheduling military aircraft and crews across missions subject to range, refueling, and personnel constraints. The solution approach combines MILP formulations with metaheuristic solvers, including Simulated Annealing, Tabu Search, and Genetic Algorithms.
Developed a stochastic programming framework that integrates product development and design decisions into supply chains facing disruptive events, informed by two Norwegian circular manufacturing case studies. Results showed resilience improvements of 4–5%, budget savings of 2–10%, and measurable gains in sustainability and circularity. Published in Omega (2025); developed during a research visit at NTNU in Norway.
Developed a three-stage matheuristic combining MILP models to solve a two-dimensional vehicle loading problem faced by a Colombian motorcycle assembler operating a heterogeneous fleet. The approach maximizes on-time deliveries, minimizes the number of trucks used, and reduces unloading rehandles during delivery stops. Across 300 test instances, the matheuristic improved the company-defined performance metric by an average of 27% over one-dimensional baselines and saved 2.69 trucks per day on average, particularly under relaxed credit constraints. Implemented in Java with CPLEX.
Modeled cooperative order picking by drones and workers in warehouses, starting from independent routing and building up to fully coordinated systems with synchronization, battery constraints, and multi-trip drone scheduling. Formulated the problem as a CVRP with MTZ subtour elimination and Big-M linearizations, minimizing makespan across multiple scenarios.
A memory-enhanced agentic system that generates Python heuristics for combinatorial optimization problems in parallel through coordinator, coding, and observer agents. The core contribution is a multi-perspective knowledge synthesis layer: five analyst LLMs, each drawing on a distinct philosophy-of-science tradition (Popperian falsification, Inductivist generalization, and three others), distill each iteration's results into evolving knowledge documents stored as episodic memory, which the coordinator reads before steering the next round of assignments. Benchmarked on Job-Shop Scheduling (OR-Library instances), single-machine tardiness, and MAX-SAT.
Built a discrete-event simulation model for a mining company evaluating proposed changes to the robots that lay out detonation grids at active sites. The model simulated grid fulfillment under different robot configurations, comparing alternatives on completion time and throughput to support the company's assessment of upgrade options.
Third dissertation chapter: trains an LSTM forecaster of vessel arrival and service times jointly with a MILP berth allocation model using decision-focused learning (SPO+ loss and perturbation-based gradients) so that the predictor minimizes schedule cost regret instead of forecast error. Evaluated against a standard predict-then-optimize baseline on Port of San Antonio (Chile) historical data, with SHAP and LIME comparing feature attributions before and after DFL training.
Research on deep learning and reinforcement learning methods for detecting patterns in illicit supply chains, with a focus on opioid precursor trafficking. Funded by the U.S. Department of Homeland Security Center for Accelerating Operational Efficiency (CAOE) in partnership with Arizona State University (2024–2025).
Built a range of portfolio and business websites, including bilingual company pages, personal sites, and this portfolio itself, each with responsive layouts and clean visual design.
Cross-platform mobile application for health and fitness tracking across iOS, Android, and the web, featuring nutrition logging, weight tracking, and progress dashboards. Started as an excuse to try AI-powered development tools (and to avoid paying for fitness apps).
A compliance advisor deployed at a Chilean pension fund administrator (AFP) to assist legal teams with financial regulations. It pairs a hybrid RAG pipeline (bi-encoder retrieval plus cross-encoder reranking over more than 4,500 legal document chunks) and multi-hop agents with a QLoRA fine-tuned Qwen2.5-7B model that answers regulatory questions and cites the source documentation.
A simulation in which LLM agents autonomously run the economy of a small town: they work, start businesses, set their own prices, socialize, and reflect at the end of each day. Business types, products, and custom actions are generated dynamically by the model, with dedicated judge agents validating the plausibility of new businesses and actions before they enter the world. Runs locally on vLLM with Qwen2.5-3B-Instruct.
Omega, Journal of Management Science
doi.org/10.1016/j.omega.2025.103317 →Two additional manuscripts are under review: a freight rate forecasting study with explainable AI (Annals of Operations Research) and a two-dimensional truck loading paper (European Journal of Operational Research).
Happy to hear from you about job opportunities, research collaborations, or just to connect on LinkedIn.