Machine Learning Engineer
I take models to production and own what happens next.
I’m Barry (Zehui) Zhang, a Machine Learning Engineer. At CVS Health/Aetna I built and ran two decision-automation systems end to end: a real-time decision service, and a daily pipeline that keeps wrongly flagged cases out of manual review. On my own time I train and evaluate deep-learning models of healthcare cost.
Tools: PyTorch · scikit-learn · Python · SQL · FastAPI · Kubernetes · Airflow · BigQuery · GCP
- Most recently
- CVS Health / AetnaMachine Learning Engineer, Feb 2025 – Nov 2026
- Education
- M.S. Biostatistics and Data ScienceWeill Cornell MedicineB.S. Joint Mathematics-EconomicsUC San Diego
- Research
- Sole author5 manuscripts in preparation, 3 patent applications pending
- Available
- ImmediatelyRemote, or relocating anywhere in the US
The whole loop, not one part of it.
- 01
Data
In production, at CVS Health/AetnaNightly pipelines that rebuild the tables the decisions are served from, a reference data product refreshed monthly, and the staging SQL that prepares each day's flagged cases.
Read the caseIn my own research72,818 person-years of national survey data, with rebuilt inputs byte-identical to the originals.
Read the research - 02
Model
In production, at CVS Health/AetnaThe decision logic of the review pipeline is a text-matching engine I implemented: pattern extraction over free-text notes, fuzzy entity matching with edit-distance limits, and a three-way answer instead of a keyword hit.
Read the caseIn my own researchA multi-task FT-Transformer in PyTorch that predicts expected cost, 7 tail quantiles and nested high-cost risk in one model.
Read the research - 03
Evaluate
In production, at CVS Health/AetnaEach new rule became a tested release, and production output was reconciled against reference runs. The monitoring dashboard was recomputed from source data before anyone relied on it.
Read the caseIn my own researchPaired bootstrap inference, calibration and a 35-run ablation. What did not reach significance is reported as such.
Read the research - 04
Ship
In production, at CVS Health/AetnaA real-time decision service on Kubernetes, released behind automated tests and a load test. A breaking contract change staged until every caller had moved. A guarded promotion path from a rule change to production that became the team's reference pattern.
Read the case - 05
Monitor
In production, at CVS Health/AetnaEvery request recorded as a structured event. Adoption, decision rate and latency by code path, published every day and reconciled to source data.
Read the case - 06
Feed back
In production, at CVS Health/AetnaNearly all missed decisions traced to specific gaps in coverage, and those went back to be closed. A healthy overall latency was hiding a slow path, so latency was split by code path.
Read the case
At CVS Health/Aetna I was the end-to-end owner of both systems and wrote most of the code in their repositories. I also owned the team's production services and pipelines across dev, test and prod.
Production systems, full products and research.
Production systems, products and research.
- CVS Health / Aetna · 2025 – 2026
Real-time decision system
A decision model, put into production as a real-time service that automation bots and an AI assistant call.
Real-timedecisions for automation bots and an AI assistant - Independent research · 2025 – present
Risk-aware healthcare cost prediction
Sole-author research on who to reach when capacity is fixed: a multi-task deep-learning model of heavy-tailed healthcare spending.
2.0×high-cost spending captured at top-0.5% capacity - CVS Health / Aetna · 2025 – 2026
Manual-review reduction pipeline
A text-matching engine and the pipeline around it, which find false-positive review flags in free-text notes and clear them every day.
Dailydelivery that keeps wrongly flagged cases out of manual review - OrchestrationAirflow 2toAirflow 3
- CI/CDJenkinstoGitHub Actions
- CloudOld environmenttoNew environment
CVS Health / Aetna · 2025 – 2026Production platform and migrations
Production services and pipelines across dev, test and prod, kept running while the platform underneath them was replaced.
Dev to prodownership of production services and pipelines, through platform migrations- Personal project · 2026
Predicting six-hour Bitcoin moves: a negative result
A blending ensemble for six-hour BTCUSDT moves, taken through a locked test and a costed backtest. The answer was no.
0.514balanced accuracy on the locked test, interval includes 0.5 
Independent project · Nov 2026Pawline: a custom pet-portrait gift store
A complete store where a customer uploads a pet photo, places it on a product and approves a designer's proof.
117automated tests across Chromium, Firefox and WebKit
Independent project · Nov 2026Wine retail commerce and operations platform
A bilingual storefront, wholesale ordering and four staff workspaces on one schema, correct under concurrency.
41tables, with role-based access for four staff roles- Resume
The whole record, on two pages
Experience, research, projects, skills and education. Read it here or take the PDF.
Machine Learning Engineer · web and PDF - This site
How this site is built and checked
The design system, the drawings, and the checks every page has to pass before it ships.
Designed, built and run by me
Figures from my time at CVS Health/Aetna are left out of this site. I am glad to walk through them in conversation.
Who should a care team reach when it can only reach a few?
Healthcare spending is heavy-tailed: a small number of people account for a large part of the total. Most cost models are tuned for the average. This research asks what a model should predict when capacity is fixed and only the top of the distribution can be acted on.
Read about the researchFlagging the top 0.5% of people, the model captures 6.24% of total spending against 3.11% for the baseline (2.01×, p = 0.018).
Count-based top-k% selection, model against a calibrated logistic baseline. Paired cluster bootstrap, B = 2,000.
An engineer with a statistician’s training.
I am a Machine Learning Engineer in Irvine, California. I interned at CVS Health/Aetna in 2024 and returned full time from 2025 to 2026, building the systems that automate decisions that used to need a human.
I studied Mathematics-Economics at UC San Diego and Biostatistics and Data Science at Weill Cornell Medicine. That mix shows up in the work: I like systems that run in production, and I like being able to say how well they run, with an interval around the answer.
- 01
Own the whole path
Design, pipelines, service, rollout, monitoring and support. The hard problems sit at the handoffs between them, so I would rather hold all of them.
- 02
Measure what changed
A system is finished when there is a number that shows it works, and that number has been checked against the source data.
- 03
Bring the callers along
A breaking change is a people problem first. I agree on the contract, stage the change and watch adoption until everyone has moved.
- ML & Modeling
- PyTorch, TensorFlow/Keras, scikit-learn, XGBoost/LightGBM, Optuna
- 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
- Observability
- Grafana, Prometheus/PromQL, custom metrics, structured logging, OpenTelemetry, load testing
- Full Stack
- TypeScript, React, Next.js, Node.js, PostgreSQL, Prisma, Playwright
- Languages
- Python, TypeScript, SQL, R, Java, C++, Bash
- Tools
- Codex, Claude Code, Cursor, Git
- Domain
- Healthcare data, RPA integration, HIPAA/PHI-safe design