Brian Plume

About Brian Plume

I design and scale reliable machine learning and LLM infrastructure that powers real-time healthcare automation.As a Senior Staff Data Scientist and ML/LLM Ops Architect at Qventus, I lead the design of multi-tenant predictive platforms used by hospitals nationwide. My focus is building systems that make data science deployable, observable, and compliant—from Databricks pipelines to Langfuse-based model telemetry.Over 15 years in data science, embedded systems, and applied AI have taught me how to bridge research and production at scale. I’ve led architecture migrations, built reusable ML frameworks, and mentored cross-functional teams spanning Data Science, Engineering, and DevOps.**Core strengths**• Databricks DLT / Unity Catalog / Delta Live Pipelines • AWS SageMaker • MLflow • Langfuse • LLM Observability • Snowflake / DBT / Feature Store Architecture • Terraform • HIPAA Compliance • Data Security & De-identification • Leadership | System Design | Cross-team Architecture My goal: deliver trustworthy AI systems that improve patient flow, reduce clinician burden, and demonstrate that applied ML can truly make healthcare work better.

Experience

1. Senior Staff Data Scientist

   Qventus, Inc

   Sep 2017 — Present · Mountain View, CA, US

   Built 40+ production ML models across ED, inpatient, preoperative, and outpatient domains, serving real-time predictions for 50+ hospital tenants.ED: census and demand forecasting (GradientBoosting with autoregressive features), likely-admit scoring (RandomForest + SelectKBest + LinearSVC pipeline), bed-ahead admission prediction, LWBS risk (KMeans text clustering with class weighting), and patient satisfaction.Inpatient: length-of-stay and GMLOS regression, multi-version estimated discharge date (EDDV3: bundled XGBClassifier + XGBRegressor with Boruta feature selection), discharge barriers with GMLOS integration, capacity management (hierarchical forecasting with LightGBM and admit/discharge/LOC reconciliation), readmission risk, severity-of-illness (multi-class with precision-recall thresholding), ICU step-down readiness, and discharge disposition.OR: surgical case-length prediction (CatBoost + GLMMEncoder), block utilization forecasting, schedule optimization (Google OR-Tools CP-SAT constraint solver), surgeon-slot matching via collaborative filtering (LightFM with WARP loss), and case volume forecasting.Consulted with Hybrid LLM + traditional: ASA physical status scoring via GPT-4o with structured clinical prompting, condition extraction and daily insights via custom fact-mining framework (LangChain + Azure OpenAI), and LLM-based inpatient explainability.Built custom ML kernels and optimizers, addressed class imbalance via SMOTE, class weighting, isotonic calibration, and time-stratified splitting. Tuning with Optuna/HyperOpt; SHAP explainability; MAPIE conformal intervals.Packages: scikit-learn, XGBoost, LightGBM, CatBoost, Optuna, SHAP, imbalanced-learn, Boruta, category_encoders, hierarchicalforecast, MAPIE, ortools, LightFM, polars, pandas, numpy, scipy, LangChain, MLflow, Langfuse

Education

• Rensselaer Polytechnic Institute

  MS, Electrical Engineering

  2002 — 2007

• Worcester Polytechnic Institute

  BS, Electrical Engineering and Computer Engineering, Minor CS

  1998 — 2002