Barry (Zehui) Zhang
Machine Learning Engineer · Production ML systems, end to end
Irvine, CAbarryzhangze@gmail.comlinkedin.com/in/barry-zehui-zhanggithub.com/BarryZ807
Public edition. Employer-specific figures are left out; a detailed resume is available on request.
Summary
Machine Learning Engineer who takes models to production and owns the systems end to end, from design to monitoring and measured impact. At CVS Health/Aetna, built and ran two production decision-automation systems, a real-time decision service and a daily review-reduction pipeline, leading delivery across business operations, Data Science and the automation team. Sole-author deep-learning research; 3 patent applications.
Experience
CVS Health / Aetna
Irvine, CA (Remote) · New York, NYDecision automation: decisions that used to need a human. Primary author of most of the team's engineering repositories.
- Put a decision model into production as a real-time service that automation bots and an AI assistant call.
- Owned the whole system from design to production support: nightly data pipelines, a privacy-safe decision API on Kubernetes released behind automated and load tests, and a lookup service on a data product I built.
- Led the rollout across business operations, Data Science and the automation team: designed the API contract, ran the working sessions and staged a breaking change until every caller had switched.
- Built the monitoring that shows whether it works: daily metrics reconciled to source data, with missed decisions traced back to gaps in the model's coverage.
- Built and ran the pipeline that finds false-positive review flags, from note matching to daily delivery to the business and automation teams, keeping those cases out of manual review.
- Implemented the matching engine (pattern extraction, fuzzy entity matching with edit-distance limits, three-way decisions) as an async Kubernetes service that Airflow starts and polls, with per-stage tracking, so one failure cannot take down the run.
- Cut runtime with async stages and tuned SQL, and shortened Data Science's path to production with an automated promotion workflow that became the team's reference pattern for new projects.
- Led the weekly sync with business operations, Data Science and the automation team, and turned each new Data Science rule into a tested production release.
Production Platform and Technical Leadership
Case study- Owned the team's production services and Airflow pipelines across dev, test and prod, and moved them through platform migrations in orchestration, CI/CD and cloud infrastructure.
- Wrote the team's engineering handbook and served as the team's bridge to a partner organization.
CVS Health / Aetna
New York, NY- Built a full-stack GenAI cost-estimation tool (front end, back end and GCP pricing data) that prices architectures such as RAG in real time, plus a recommender that proposes options within budget.
Institute for Community Living (ICL)
New York, NY- Built auto-refreshing Power BI dashboards used across departments and quarterly KPI reports in Python, SQL and R.
Research
- Designed a multi-task FT-Transformer (PyTorch) that predicts expected cost, cost quantiles and high-cost risk from 72.8K person-years of US national survey data, with 0 quantile crossings vs. 1.54% for baselines.
- Doubled the high-cost spending captured vs. a calibrated baseline when flagging the top 0.5% (2.0x, p = 0.018), with significant gains across the top 0.5–2% and good calibration (ECE 0.006).
- Built to research-grade standards: 60 seeded, reproducible runs with 0 failed steps, survey-weighted bootstrap inference, and all 98 reported numbers machine-verified against run outputs.
Independent projects
Wine Retail Commerce & Operations Platform
Case studyPortfolio build modeled on a family-owned wine retailer's public catalog · Next.js, React, TypeScript, PostgreSQL
- Designed and built the full platform: a bilingual storefront, wholesale ordering, and owner, advisor, warehouse and support workspaces with role-based access on a 41-table PostgreSQL schema.
Pawline: Custom Pet-Portrait Gift Store
Case studyOwn brand concept, pre-launch · Next.js, React, TypeScript, PostgreSQL, Playwright
- Verified it like production: 117 automated tests with Chromium, Firefox and WebKit runs, a 25.9K-request load test (read p95 20–53 ms), exactly 120 charges for 120 concurrent orders and forced-crash recovery.
Skills
- ML & Modeling
- PyTorch, TensorFlow/Keras, scikit-learn, XGBoost/LightGBM, Optuna; tabular deep learning (FT-Transformer), multi-task heads, quantile regression, calibration, ablation, decision-curve analysis
- Serving & MLOps
- FastAPI, Pydantic, Docker, Kubernetes (GKE), REST API design, streaming, rate limiting, GitHub Actions, Jenkins, Snyk, SonarQube, pytest, HashiCorp Vault
- Data & Cloud
- GCP (BigQuery, Airflow on Cloud Composer 2 and 3, GKE, GCS, Cloud Monitoring/Logging, Secret Manager, IAM), AWS; SQL, pandas, NumPy
- Observability
- Grafana, Prometheus/PromQL, custom metrics, structured logging, OpenTelemetry, load testing
- Full Stack
- TypeScript, React, Next.js, Node.js, PostgreSQL, Prisma, Playwright
- Languages & Domain
- Python, TypeScript, SQL, R, Java, C++, Bash; Codex, Claude Code, Cursor, Git; healthcare data, RPA integration, HIPAA/PHI-safe design
Education
Weill Cornell Medicine, Cornell University
New York, NYCoursework: AI in Medicine, Computer Vision, Biostatistics I & II, Categorical & Censored Data Analysis, Statistical Learning
University of California, San Diego
San Diego, CACoursework: Intro to Machine Learning, Data Structures & Algorithms (Java), Data Management, Probability, Mathematical Statistics I & II
Manuscripts & patent applications
Lead manuscript: Healthcare cost prediction when capacity is fixed: one model for expected cost, tail quantiles and high-cost risk (prepared for submission to JAMIA); 4 companion manuscripts in preparation.
Patent applications (China, CNIPA; filed 2026, pending)
- A method for predicting healthcare spending.
- A method for grouping a population by predicted risk.
- A method for allocating care resources from predicted risk.