Randolph Schilke, MPH

• Architecting MLOps v3, an enterprise hub-and-spoke platform on AWS: a hub of shared, KMS-encrypted S3 data tiers with cross-account trust, the coe-utils library of reusable CDK constructs, and a Copier template that generates ready-to-deploy ML projects wired to the hub, so onboarding a new model needs no platform change.
• Designed registry-first model promotion: a lifecycle change in the dev SageMaker Model Registry triggers an EventBridge-driven orchestrator and cross-account promotion pipelines with a forward-only gate, human approval for prod, and idempotent registration, so models move between environments without redeploying code.
• Building multi-region resilience: cross-region replication of approved model artifacts, an event-driven registry mirror with scheduled reconciliation, and secondary-region pipelines that wait on the primary.
• Architected and own CarePath, a multi-account AWS MLOps platform running 37 models across 8 lines of business (Medicare Advantage, Commercial, M360, Maternity, Behavioral Health, Condition Care), with all infrastructure defined in Python CDK.
• Built 3-stage cross-account CI/CD promotion (dev → stage → prod) on CodePipeline with CodeStar Connections to GitHub Enterprise, branch-specific triggers, and deterministic CloudFormation asset publishing via CodeBuild.
• Designed cross-account model promotion: KMS-encrypted silver-to-gold artifact buckets with cross-account S3/KMS policies; a Lambda custom resource validates artifacts and auto-registers and approves versions in SageMaker Model Registry on every deploy.
• Built SageMaker training pipelines running cohort creation → training → parallel baseline generation (Data Quality, Model Quality, Bias) via Model Monitor and Clarify, with version-stamped artifacts for reproducibility.
• Implemented cron-scheduled data quality and bias monitoring through SageMaker MonitoringSchedules at the model-family level, with CloudWatch alarms and dashboards.
• Stood up a separate PR validation pipeline running pytest and linting in VPC-connected CodeBuild, plus an SNS and Lambda construct that posts commit status back to the GitHub Enterprise API.

stack AWS CDK (Python), CDK Pipelines, Copier, SageMaker Pipelines / Model Registry / Model Monitor / Clarify, EventBridge, Lambda, SQS, CodePipeline, CodeBuild, CloudFormation, S3 (cross-region replication), KMS, IAM, VPC, SSM, Secrets Manager, SNS, CloudWatch, ECR, GitHub Enterprise, Snowflake

• Built the VA’s first production LLM pipeline: an end-to-end Azure Databricks pipeline for clinical note processing. Authored the abstract, presented to VA stakeholders twice to secure approval, and left it running in automation with an estimated $5M+ in annual savings.
• Maintained and extended the 30-Day Decompensation program, the longest continuously running ML pipeline in VA history.
• Delivered the MIRTH program’s infrastructure end-to-end in AWS CDK (Python): S3, Lambda, Step Functions, SageMaker Processing, ECR, SNS, IAM, and CloudWatch, with updated architecture documentation for repeatable deployment.
• Authored the VA’s first external API data ingestion integration, pulling Apple App Store and Google Play data into HDAP; still running in automation.
• Unblocked a TeleDerm computer vision project stalled for a year by labeling 10,000+ dermatology images and building a reusable annotation application in AWS.
• Established data science standards for the OCC: sandbox repository, MLflow experiment and registry conventions, GitHub repositories for all OCC projects, and written onboarding guides. Led code reviews and mentored the NAII team.

stack Azure Databricks, Spark, MLflow, AWS SageMaker, CDK, Lambda, Step Functions, GovCloud (AWS/Azure/GCP), VA GitHub Enterprise

• Golf Ball Tracer (2026): broadcast-quality golf shot tracing for iOS and Android.
• Chores Hero (2026): family chore management app for iOS and Android.
• 3LYTRA (2025–2026): autonomous robot platform with WebRTC remote control (aiortc, Firebase RTDB signaling, TURN NAT traversal), on-device LLM inference via Dockerized llama.cpp with CUDA on Jetson Orin Nano, YOLO object detection on Raspberry Pi 5, and BLE WiFi onboarding.
• Nuclear Winter (2024): mobile survival game driven by an agentic, hierarchical LLM pipeline (OpenAI, LangChain/Langflow, crewAI) in Dockerized microservices on AWS App Runner and ECS Fargate, with Amplify (Cognito) and Firebase auth.
• BeachBop (2023): 18 on-device Core ML classifiers trained on 100,000+ labeled images, ARKit/LiDAR tooth measurement, and a 2,500-photo reference gallery across 83 species and genera. USPTO trademark #7699228.

stack React Native, Swift/SwiftUI, Core ML, ARKit, Firebase, AWS Amplify, App Runner, ECS Fargate, llama.cpp, CUDA, YOLO, Raspberry Pi 5, Jetson Orin Nano

• Managed a 10-person data science team of senior and lead individual contributors; facilitated round tables, think tanks, 1:1s, and stakeholder reviews.
• Established the organization’s first MLOps pipeline in Azure Databricks (MLflow, PySpark) with production scripts for feature store, inference, retraining, and validation; built Mercy’s first real-time models integrated with EPIC EMR.
• Productionalized four models through Databricks Workflows (Mercy Works on Demand, Customer Retention, Clinic Volume Forecast, Gross Financial Forecast) and migrated the team from individual notebooks to Azure DevOps version control with mandatory peer review.
• Applied John Snow Labs embeddings and Azure Text Analytics for Health to structure unstructured clinical text; built a Power BI model validation dashboard on Azure Synapse.

stack Azure Databricks, MLflow, PySpark, Delta Tables, ADLS, Synapse Analytics, Text Analytics for Health, Azure DevOps, EPIC EMR, Power BI

• Held production oversight of 45 active models: deployment, debugging, redeployment, and post-production monitoring, including IT change request management and peer code review.
• Built propensity-to-buy models for Medicare enrollment (O65 AEP South and Non-South Florida, U65 OEP subsidy and non-subsidy), care management engageability, Medicare Advantage disenrollment, precision health management, and the 2022 Federal Employee Program model.
• Developed NLP and sentiment analysis pipelines for business segment applications; discovered data pipeline leaks and deployed system patches.
• 1st place, 2020 GuideWell Data Science Hackathon; Honorable Winner, 2021 GuideWell Innovation Challenge. Managed and mentored four interns and junior data scientists.

stack Python, H2O, gradient boosting, logistic regression, NLP, SQL Server, DB2, Jenkins, GitLab

• Automated state- and county-level quarterly non-fatal overdose reporting and SUDORS fatal case uploads in SAS; authored seven public-facing state surveillance reports.
• Applied text mining to EMS clinical notes to improve syndromic surveillance case definitions and predictive models; presented findings to the CDC EMS All-State Workgroup and the Council of State and Territorial Epidemiologists.
• Identified geographically specific, statistically significant changes in non-fatal overdose rates using Florida ESOOS EMS data.

stack SAS, SAS Enterprise Miner, SAS Text Miner, SQL, Python

• Managed the Transplant Institute data warehouse and delivered enterprise analysis across UNOS STAR, SRTR registries, and EPIC EMR.
• Built survival models and propensity score matching pipelines supporting peer-reviewed publications in Gastroenterology, Hepatology, and Transplant International.
• Presented clinical research at the American Transplant Congress and the International Congress of The Transplantation Society; mentored junior analysts.

stack SAS, SQL, Python, EPIC EMR, UNOS STAR, SRTR