Work Experience

Data Scientist — Elevance Health, Inc. (formerly Anthem Inc.)

March 2021 – Present  ·  Chicago, IL

I've been building and owning production ML and LLM systems to automate clinical prior authorization — a process that traditionally requires physicians and nurses to manually review patient records against clinical guidelines. My work spans 30+ medical guidelines and multiple business units, and has collectively saved tens of thousands of clinical hours annually.

A large part of my recent focus has been on large language models. I designed a DAG-based LLM pipeline with swappable processing nodes, Pydantic-enforced structured outputs, and robust retry logic, first deployed for therapy visit authorization, where it achieves 90% accuracy and generates around $200K in monthly savings. At its core is a hybrid RAG system combining BM25 sparse retrieval with OpenAI dense embeddings via Qdrant, plus cross-encoder reranking, which cut context token usage by 40% while improving clinical document search quality. The core pipeline is now used by multiple workstreams.

I also built a production-grade vision-language model (VLM) pipeline powered by GPT-5.4 that extracts behavioral assessment scores from 500K+ clinical documents, reaching 95–99% accuracy across six different assessments and cutting latency by 90%. The pipeline handles image resizing, request batching, and smart retries that adapt to API rate limits. Its outputs now serve as core quality metrics for downstream anomaly detection.

Alongside the LLM work, I've trained and maintained LightGBM models across 15+ guidelines, tuning decision thresholds against target false-positive rates and applying SHAP values for clinical explainability — enabling 30% of cases to be automatically approved and saving roughly $100K per month. This work eventually led to a US patent (US 12293835 B1) for a novel AI-based authorization automation system, and a second patent application is currently being filed. I also built clinical NLP pipelines using UmlsBERT and CODER embeddings with spaCy and Stanza for named entity recognition, extracting structured medical facts from free-text clinical notes at 85% precision.

On the engineering and team side, I contributed to the MLOps foundation by building pip-installable internal packages with UV and Poetry, integrating GitLab CI/CD pipelines for linting, testing, and deployment, and cutting release cycle time by 50%. I also led a pod of six data scientists, architecting a scalable, reusable codebase that could be rapidly adapted across new clinical guidelines — driving over 25 model deployments in five months.

Machine Learning Consultant — Hamiltonian Systems, Inc.

June 2020 – December 2020  ·  Pittsburgh, PA

I built machine learning models to predict weekly transaction quantities and annual transaction occurrences, using Keras and TensorFlow to construct neural networks trained on historical usage data, achieving over 70% accuracy. I also led a team of five to build inventory lifecycle prediction models using Random Forest and XGBoost on 440K+ transactions, reducing forecast error by 22% over baseline.

After the models were ready, I designed a multithreaded Flask API to orchestrate ML model training and inference pipelines with async task management and cloud deployment via REST endpoints — reducing manual engineering workload by 90% and making the entire pipeline accessible with a single button click.

Research Assistant — Heinz College, Carnegie Mellon University

March 2020 – May 2020  ·  Pittsburgh, PA

As a Research Assistant at Carnegie Mellon University, I collaborated with Min-heng Wang on Hope Hurts: How High Expectations on Special Days Disappoint Yelp Users. The research investigated how elevated expectations around special occasions — birthdays, anniversaries, holidays — shape the way people experience and review restaurants.

To answer that question, I analyzed 163,000 Yelp reviews using Structural Topic Modeling in R to uncover latent sentiment patterns and topic clusters. I then applied regression analysis alongside Propensity Score Matching to control for selection bias and estimate the effect of special-day dining on user satisfaction scores.

Consultant Intern — Deloitte Consulting

November 2018 – May 2019  ·  Shanghai, China

Data Analyst Intern — Dell Technologies

July 2018 – October 2018  ·  Shanghai, China

Business Development Intern — Infralutions, Inc.

January 2018 – February 2018  ·  Hong Kong, China (remote)