Washington, District of Columbia, United States
I specialize in developing AI-driven optimization and machine learning solutions to drive efficiency and impact at scale. With over eight years of experience, I’ve worked on building predictive models, optimization frameworks, and scalable ML pipelines that enhance decision-making in complex systems. 🔹 Optimization & AI – Designed and deployed advanced models for large-scale decision-making, leading to significant cost savings and sustainability improvements. 🔹 Machine Learning & Forecasting – Leveraged deep learning and predictive modeling to improve forecasting accuracy across diverse applications. 🔹 Scalable ML Systems – Built and deployed end-to-end ML pipelines using cloud platforms, ensuring robust and efficient model performance. 🔹 Leadership & Strategy – Lead a cross-functional data science team, driving best practices in AI, optimization, and MLOps to create real-world impact. Always exploring innovative ways to apply AI and optimization for smarter decision-making. Let’s connect!
• Leading the design, development, and deployment of supply chain network Digital Twin optimization models using mixed-integer programming (MIP) in Python, OR-Tools, and Gurobipy, minimizing costs, lead times, risks, and emissions; achieved $202M in EBIT and reduced CO₂e emissions by 100M kg across 1.6 billion units of apparel and accessories • Developing and maintaining multilayer perceptron (MLP) models in PyTorch to predict apparel manufacturing and tariff costs using product attributes; integrating cost predictions into sourcing optimization models • Building and supporting CatBoost models to predict footwear component costs and manufacturing time (minutes) based on product attributes; delivering results via a Streamlit dashboard used by 150+ designers to accelerate the design process • Designing and optimizing large-scale ETL pipelines in AWS SageMaker and Databricks using the medallion architecture, automating ingestion from sources including Snowflake, Box, and Databricks Delta Tables; containerizing pipelines with Docker and implementing PyTest for robust testing and deployment • Building a Retrieval-Augmented Generation (RAG) system with AWS Bedrock and LangChain for internal documentation retrieval; using Databricks Genie to explain optimization results, improving stakeholder understanding and decision-making • Mentoring data scientists, engineers, and interns on best practices in optimization modeling, ML workflows, and production-grade code to improve model quality and deployment efficiency
• Developed a two-stage optimization model using MIP in Python, PySpark, and PuLP on Palantir Foundry for sourcing and delivery scheduling, and container packing to minimize shipments while meeting demand and inventory constraints • Developed an LSTM-based sequence-to-sequence model in TensorFlow to forecast 12-month backorders across 30K SKUs using five years of time-series data; achieved 0.89 F1-score and 0.94 recall via walk-forward cross-validation • Implemented Multilayer Perceptron and Recurrent Neural Network (RNN) models using scikit-learn and TensorFlow to predict the reliability of aircraft engine components from unstructured text, classifying 30+ modes of failure with 85% accuracy • Engineered data pipelines in Pandas, PySpark, and SQL to support ML forecasting and network optimization workflows, including time-series feature engineering, scenario input processing, and ingestion of product metadata, transit routes, and inventory levels; enabled UI-driven model execution and results visualization • Integrated SHAP and LIME into model pipelines to visualize feature importance and explain predictions from ML models • Used tslearn to apply time-series k-means clustering for demand pattern segmentation and cluster-guided oversampling, improving data balance and enhancing backorder forecast accuracy across underrepresented product groups
• Designed and implemented temporal U-Net neural network models to generate segmentation maps of solar surface phenomena from satellite imagery, achieving a 106% accuracy improvement over the previous model. Developed in Keras with a TensorFlow backend, trained on dual GTX 1080 Ti GPUs. • Built an end-to-end preprocessing pipeline in Python (scikit-learn, NumPy, pandas) to clean and prepare 30,000+ observations (~300 GB) from NASA’s Solar Dynamics Observatory, including luminosity correction, image augmentation, and data normalization. • Conducted literature reviews on state-of-the-art computer vision models (CNNs, RNNs) for spatiotemporal pattern recognition, tuned hyperparameters, and integrated findings into model development. • Produced technical briefings with visualizations in matplotlib, and collaborated with senior data scientists via GitLab to iterate on model architecture and improve deployment readiness.
• Developed and implemented dual input recursive neural network models to classify hate speech in Wikipedia and Twitter comments via natural language processing, TensorFlow, and Keras using an RTX 2070 GPU and increased performance by 0.65% over baseline models • Generated sentence embeddings using the Google Universal Sentence Encoder and bias via Word Embedding Factual Association Test (WEFAT) association scores from GloVe word embeddings using Python (NumPy, pandas, and scikit-learn) • Performed literature review on state-of-the-art hate speech detection methods and produced technical report to document methods and results
• Built a SKU recommendation system in SAS and SQL, sourcing data from Oracle and clustering products by vendor and item attributes; pushed outputs to a VBA dashboard used by buyers to explore and finalize contract-ready groupings • Led development of a supply chain network dashboard in Qlik Sense and engineered data pipelines in Python and SQL • Conducted ad-hoc analyses on recommendation engine group quality by querying Oracle and SAP HANA databases • Automated manual processes by developing Python and VBA to generate email content and financial reports, reducing errors, saving approximately 5 hours per month, and improving overall report accuracy