Ph.D. in Computational Statistics
San Diego State University / Claremont Graduate University, 2018
Resume
Machine learning systems from raw data to production.
Senior machine learning engineer with experience across modeling, distributed data systems, deployment, reliability, applied AI, research, and teaching.
Experience
Consulting Research Engineer, Climate Data Platform SDSU
- Built reusable climate-data pipelines to process heterogeneous scientific datasets and publish outputs directly into databases, Parquet, and Zarr for downstream analysis and platform use.
- Worked across geospatial and scientific data formats including Zarr, NetCDF, GeoTIFF, and shapefiles to support ingestion, transformation, and publication workflows for climate datasets.
- Designed the processing workflows to be reusable across different climate datasets rather than one-off scripts, supporting repeatable ingestion and transformation patterns for the platform.
- Built and maintained platform-oriented climate-data processing workflows for the iCHARM website and related SDSU research use cases.
Senior Machine Learning Engineer Unit21
- Built AI-agent tasks that customers used to automate time-consuming workflows.
- Exported selected Databricks datasets to versioned Parquet files in S3, enabling faster metric calculation with DuckDB and preserving reproducible audit trails for statistical analysis outputs.
- Designed and built ML systems for AML and fraud detection using heterogeneous, customer-specific datasets across hundreds of customer environments.
- Built automated feature generation pipelines for customer-provided custom data, enabling substantial model performance gains for customers with rich, predictive data sources.
- Improved model performance in high-lift customer environments from approximately 0.60 AUC to as high as 0.97 AUC through automated feature engineering and model pipeline improvements.
- Re-architected large-scale Spark processing from Spark SQL to optimized RDD-based workflows, reducing runtime from approximately 12 hours to roughly 30 minutes on the same cluster.
- Built an LLM-assisted model explainability workflow that mapped per-observation SHAP values to raw feature values, feature metadata, and model scores to generate clear score narratives.
- Built AI-agent workflows using AWS Bedrock for financial investigation, data analysis, and customer-defined tasks.
- Designed FastAPI-based model execution services integrating distributed feature generation, model scoring, and production-facing prediction workflows.
- Led ML and AI research prototypes that informed or evolved into production-facing platform features.
- Researched consortium-based fraud intelligence sharing approaches to improve cross-customer risk detection and bad-actor identification.
- Cleaned and standardized customer-specific datasets with inconsistent schemas, formats, and data quality issues into reliable, model-ready inputs for machine learning.
- Migrated ML processing workflows from Snowflake-based execution to PySpark/Databricks, improving scalability, maintainability, production performance, and control over large-scale analytical computation.
- Designed and maintained Nix/nix-darwin infrastructure for macOS engineering environments, standardizing dependencies and improving reproducibility across engineering teams.
- Participated in weekly on-call rotations for production ML and data systems, owning incident response and reliability for customer-facing platform workflows.
- Developed a SQL lineage tool to identify required model features and dependencies, automatically pruning unused scripts and feature logic to dramatically speed up model data pipelines.
- Automated local development startup workflows to make local testing faster and easier.
Adjunct Professor – Machine Learning Engineering SDSU
- Developed and taught Machine Learning Engineering coursework focused on practical ML systems, model deployment, production workflows, and applied engineering patterns.
- Created public lectures, slides, and examples used by students learning modern ML engineering concepts.
Senior Machine Learning Engineer AppFolio
- Built production ML monitoring dashboards in Redash to track accuracy trends, drift, customer-specific performance, and overall model health across deployed models using real-world invoice-resolution feedback.
- Improved a BI data consolidation tool that merged sharded customer databases into a single Snowflake analytics database using Debezium and Python, resolving schema drift, data-type conflicts, and warehouse-specific merge/upsert issues across rolling deployments on Kubernetes.
- Built an end-to-end Spark-based ML training and deployment platform that used custom RDD tasks to train 1,400 customer-specific Scikit-learn invoice prediction models in parallel on a cluster.
- Built a custom Airflow-on-Kubernetes service using Helm charts, with cluster provisioning, permissions, and related infrastructure scripted in Terraform.
- Built Airflow DAG workflows that ran custom Docker images on Kubernetes clusters to automatically train and update customer models.
- Built a fully automated ML training and deployment platform supporting approximately 1,400 customer-specific invoice prediction models.
- Designed cloud-native ML infrastructure using Kubernetes, Docker, Airflow, Terraform, and production deployment workflows.
- Developed document understanding models to extract invoice metadata including vendor, invoice amount, invoice date, and related fields from customer invoices.
- Built conversational AI and customer-support chatbot systems using BERT-based NLP models, custom feature engineering, intent prediction, response selection, and next-question recommendation workflows.
- Converted ML models to AWS Lambda containers for both cost savings and faster deployments.
- Implemented CI/CD workflows for automating data preparation, model training, validation, deployment, and production updates.
- Improved existing machine learning models through feature engineering and model selection.
- Used rapid invoice-resolution feedback to compare model behavior and evaluate candidate improvements across deployed customer-specific prediction systems.
- Published trained Scikit-learn invoice prediction models to cloud storage for production serving by application services.
- Participated in weekly on-call rotations for production ML systems and platform infrastructure, owning incident response and operational reliability for deployed customer models.
Data Scientist Experian
- Re-architected Hadoop-based data processing workflows to Apache Spark, reducing processing time and AWS infrastructure cost.
- Designed and implemented an end-to-end AWS EMR optimization system for cookie-replacement and digital identity models, evaluating candidate variables against entropy, longevity, and validation-performance metrics in a survival-analysis-style framework for offline model comparison.
- Built automated reporting workflows using SQL, R, R Markdown, and Shiny to summarize candidate model performance, compare feature sets, and support statistical analysis of digital identity models.
- Developed distributed data systems using Hadoop, Pig, Java, SQL, Spark, and Airflow for feature analysis, model evaluation, and reporting.
- Designed optimization workflows for evaluating large combinatorial modeling spaces that were computationally infeasible to search exhaustively, supporting large-scale offline model and feature comparison.
- Created Airflow-based orchestration for local and AWS-based data pipelines, from ingestion through modeling and final reports.
- Expanded available predictive variables from approximately 100 to over 2,000 through automated feature engineering and data analysis.
- Built a Hadoop/Spark cluster from scratch using decommissioned hardware when the team lacked a dedicated environment for Spark experimentation.
Statistical Modeler Digital Risk
- Built end-to-end predictive modeling systems using credit, property, CSV, web, API, and third-party data sources to prioritize loans for forensic audit and defect detection.
- Designed and maintained a SQL Server research database with over 9 million observations and more than 5,500 candidate predictive variables.
- Developed mortgage, valuation, and financial risk models using R, SAS, Matlab, SQL Server, C#, and statistical modeling workflows for high-stakes forensic-audit decision support.
- Implemented a Bayesian network in C#.
- Built machine learning automated valuation models (AVMs) for real estate assets, including apartment-complex-level models using property, credit, geographic, and third-party data sources.
- Engineered geospatial valuation features, including shapefile-based distance-to-coast calculations for Florida properties where proximity materially impacted property value.
- Generated 10-year real estate investment analysis reports estimating property income, expenses, and projected future property value.
- Wrote multithreaded optimization tools in Matlab and C# to accelerate variable selection and modeling workflows, improving one process by roughly 20x over the prior non-optimized approach.
- Implemented production scoring systems and automated internal and customer-facing reports using SQL Server, SSRS, and C#-generated PDFs with maps, visualizations, charts, and analytical summaries for audit-targeting workflows.
- Solved workforce distribution optimization problems to increase employee productivity.
- Compared baseline and candidate models, including head-to-head testing of competing approaches, to improve statistical targeting of loans for forensic audit.
Doctoral Researcher, Computational Statistics SDSU / CGU
- Developed spherical-harmonic statistical methods for dimension reduction and global climate signal extraction from high-dimensional geospatial data on the surface of a sphere.
- Stabilized rank-deficient matrix and eigensystem calculations using diagonal-perturbation and regularization techniques in repeated large-scale harmonic-specific statistical workflows.
- Built a hybrid Java and C/C++ numerical pipeline, using JNI-based Java wrappers around native code to accelerate computationally intensive matrix and estimation routines; used Matlab for visualization and plotting.
- Re-architected the research computation to run on Spark by modeling harmonic-specific workloads as distributed classes/RDD tasks, making thousands of expensive regularized computations tractable across a cluster.
.NET Application Developer / QA / Technical Support CoreLogic
- Built an end-to-end credit card bust-out fraud model for a major banking institution under a one-month deadline, covering data cleaning, feature extraction, model development, evaluation, and delivery in a form the client could implement with minimal system changes.
- Validated both incoming client data integrity and the statistical methodology used in production code, helping ensure production model correctness.
- Built a survival-analysis / net-present-value optimization model for mortgage-loss mitigation, comparing foreclosure, short-sale, and loan-modification scenarios under varying bounded input parameters for decision support.
- Rewrote an existing numerical optimizer in C#, making it roughly 10x faster through multithreading, mathematical simplification, and formula-level refactoring.
- Built fraud detection, mortgage-risk, and predictive analytics machine learning models for financial services clients.
- Developed C# production code for mathematical, statistical, and optimization components of the WillCap mortgage-risk project.
Education
M.S. in Statistics
San Diego State University, 2014
B.S. / B.A. in Statistics / Mathematics
San Diego State University, 2009
Publications
Pierret, J., and S.S.P. Shen, 2017. 4D visual delivery of big climate data: A fast web database application system. Advances in Data Science and Adaptive Analysis.
Shen, S.S.P., Pierret, J., et al., 2020. 4DVD visualization and delivery of 20th century reanalysis data: methods and examples. Theoretical and Applied Climatology.