kevin liu

Sr AI/ML Engineer

Lyra Health

Mar 2024 - Current (2 years 2 months)

Summary: Led the development of HIPAA-compliant, LLM-powered systems to support mental health triage, clinical summarization, and patient support workflows. • Led the design and deployment of full-stack RAG pipelines using GPT-4, Qdrant, and LangChain, reducing support resolution time by 32%. • Mentored junior ML engineers and new hires on GenAI tooling, best practices, and evaluation methodology. • Fine-tuned LLMs for triage and Q&A tasks using PEFT and LoRA. • Deployed asynchronous LLM inference routing using Ray Serve across multiple models with load balancing. • Used Ray Tune for distributed hyperparameter tuning on transformer-based models. • Built scalable data pipelines with Databricks and Delta Lake to prepare multi-source hea

G

Sr AI/ML Engineer

GenAI Systems

Feb 2022 - Feb 2024 (2 years 1 month)

Summary: Supported Claude LLM product experimentation and RAG benchmarks while advising enterprise clients on GenAI system integration. • Implemented Prometheus + Grafana dashboards to monitor benchmark runs, GPU utilization, and response latency in RAG pipelines. • Contributed to RAG system design for enterprise clients using pgvector, Qdrant, and function-calling LLM agents. • Prototyped hybrid search strategies and safety filtering pipelines in LangChain and Haystack. • Advised engineering teams on Ray Serve deployments and caching strategies for high-availability inference. • Evaluated hybrid GenAI stacks with GCP Vertex AI, OpenSearch, and Ray Serve, comparing latency, cost, and scaling trade-offs. • Explored multi-modal RAG use

S

Sr AI/ML Engineer

Jul 2017 - Jan 2022 (4 years 7 months)

Summary: Designed and deployed scalable, production-grade machine learning systems on AWS infrastructure, with deep contributions across GenAI, computer vision, fraud detection, and MLOps. Acted as a technical bridge between AWS solutions teams and enterprise clients to accelerate adoption of AI/ML tools and architectures. • Developed distributed training pipelines using Databricks on AWS to process 100M+ record fraud datasets, enabling real-time risk scoring and dynamic rule evaluation. • Prototyped LLM-serving stack using Ray Serve + Triton Inference Server for GenAI demos. • Created GenAI architecture guides integrating LangChain, Qdrant, and OpenSearch. • Built SageMaker Pipelines for end-to-end model development lifecycles, coverin

Graduate Research Assistant / ML Engineer

Southern Methodist University

Sep 2015 - May 2017 (1 year 9 months)

Summary: Worked on applied machine learning projects in NLP and computer vision while pursuing a part-time graduate degree in Computer Science. • Developed classification and entity recognition models using scikit-learn and spaCy. • Built OCR pipelines using OpenCV and Tesseract to process scanned documents. • Created interactive dashboards for faculty using Jupyter, pandas, and seaborn. • Supported PyTorch-based training of text classification models on GPU clusters. • Maintained version-controlled notebooks and experiment logs for reproducibility. • Assisted with data anonymization tasks for clinical text projects. • Built early FastAPI endpoints for internal model testing. • Contributed to internal documentation and research tool

Data Scientist

Stripe

Jun 2013 - Aug 2015 (2 years 3 months)

Summary: Developed data-driven fraud detection and risk models to support Stripe Radar, and contributed to productionizing real-time ML scoring systems. • Designed fraud detection models using XGBoost and LightGBM, improving precision by 22%. • Created ETL pipelines in Airflow + Spark + Snowflake for fraud signal generation. • Built real-time scoring endpoints using FastAPI and Kubernetes with <100ms latency. • Collaborated with risk analysts to implement dynamic model thresholds and manual review queues. • Built internal dashboards in Streamlit and Tableau to track fraud trends and drift. • Automated training data refresh pipelines for 5+ regional markets. • Introduced early use of MLflow for model tracking. • Benchmarked embedding

M

Data Science Intern

Microsoft Research (Health Intelligence Group)

Jun 2012 - May 2013 (1 year)

• Built scikit-learn-based classifiers for labeling biomedical abstracts. • Conducted exploratory analysis on term distributions and TF-IDF features. • Helped evaluate reranking strategies for question-answer retrieval systems. • Contributed to data annotation scripts for internal training datasets. • Supported literature review and feature design for baseline NLP tasks.

About

Detail

Timeline

Résumé