top of page

AI/ML & MLOps: From Clinical RAG to Production Forecasting

Jul 23
2 min read

Over the last seven years my work has shifted from pure firmware toward machine learning that has to survive contact with production: sensitive data, reproducibility requirements, and models that keep working after they ship. A few projects that capture that.

ReliSource Protocol Parser — a private clinical-trial RAG platform

A local-first GenAI/RAG platform for clinical-trial protocol intelligence. The hard constraint was privacy: clinical protocol data can't leave a controlled environment, so the entire extraction and retrieval stack had to run without shipping sensitive text to third-party services.

  • Built extraction on hosted and local LLMs with embeddings and hybrid vector search, so reviewers can query protocols in natural language and get cited answers.

  • Layered human validation, source citations, structured exports, and mobile-ready search workflows on top of the retrieval layer.

  • Kept all sensitive clinical data inside a private environment while still cutting manual protocol-review effort.

  • Instrumented the pipeline with LangChain and LangSmith for traceability and evaluation.

Cargo demand forecasting on Azure ML

An end-to-end MLOps pipeline for cargo demand forecasting, built for reproducibility rather than one-off notebooks.

  • Automated training and deployment with Azure ML and Azure DevOps pipelines, shipping models to Azure Container Instances and Kubernetes.

  • Added a model registry, CI/CD automation, drift monitoring, and traceable retraining so every production model can be reproduced and explained.

  • Improved forecasting accuracy and operational efficiency for demand planning.

Medical and diagnostic models

  • Rheumatoid arthritis: Built an adaptive deep neural network (TensorFlow/Keras) to predict rheumatoid arthritis disease activity, supporting personalized treatment planning.

  • Spinal abnormalities: Developed a Python model to identify spinal abnormalities from biomechanical angles, improving diagnostic accuracy for clinical decisions.

  • Contributed to AI-driven diagnostics for chest diseases as part of the same clinical-ML effort.

Data science for public policy

  • Modeled US high-school dropout rates and built a Python + Plotly dashboard that surfaced the biggest contributing factors and turned them into actionable insight for policymakers.

  • Predicted US state-level student scores from demographic and financial factors to support data-driven education strategy.

  • Designed a scalable web crawler (Selenium + BeautifulSoup) that scraped and organized 15,339 municipality images across all 50 US states.

Leading the team

Directed a 12-engineer multidisciplinary team across GenAI, full-stack, and QA, accelerating sprint velocity by 23% through agile cross-team workflows.

Tech stack: Python, LangChain, LangSmith, embeddings, vector databases, hybrid search, TensorFlow/Keras, scikit-learn, Optuna, Azure ML, Azure DevOps, Kubernetes, Docker, Flask, Plotly.

 
 
 

Comments

bottom of page