Premier Analytics Consulting at the Southeast and Atlantic SAS® Users Group (SEASUG) 2026 Conference

This paper presents a practical framework for building reliable, reproducible, and collaborative data science workflows using Git, GitHub, Data Version Control (DVC), and MLflow. It begins by introducing Git as the foundation for tracking code changes and GitHub as a collaboration layer for shared repositories, branching strategies, pull requests, and team-based development. DVC is then presented as an extension to version control that enables datasets, intermediate outputs, and analytical artifacts to be tracked, compared, and restored alongside code without storing large files directly in Git repositories. The paper demonstrates strategies and techniques for meaningful commit practices, managing work-in-progress (WIP) safely, reducing merge conflicts, and maintaining structured project histories across collaborative environments. Building on this foundation, MLflow is introduced as a lightweight experiment tracking and model management layer for machine learning workflows in Python, allowing teams to record training runs, fine-tuning parameters, evaluation metrics, model artifacts, and performance comparisons during iterative development. Together, Git, GitHub, DVC, and MLflow provide an integrated open-source ecosystem for managing code repositories, versioning data, and tracking model experiments across collaborative data science, machine learning, and analytics projects.

Artificial Intelligence (AI) continues to reshape business, technology, science, and research by enabling systems to learn from data, automate workflows, and support more adaptive decision-making. This paper presents a practical roadmap for understanding AI through the progression from machine learning (ML) foundations to modern generative and agentic AI systems. It introduces AI as a broad field and machine learning as a data-driven approach to AI, then examines supervised learning for predictive analytics on labeled data, unsupervised learning for pattern discovery in unlabeled data, and generative learning for representation learning, synthesis, reasoning, and content generation. The paper introduces key algorithms, architectures, use cases, model hyperparameters, and practical considerations such as overfitting, underfitting, and model evaluation to support applied implementation across industry settings. It concludes by connecting deep learning with modern generative model architectures, including encoder models, decoder-only large language models, and encoder-decoder models, to the emerging shift toward agentic AI systems that combine models with retrieval, tools, APIs, and workflow orchestration to support practical industry use cases in 2026, 2027, and beyond.

Generative AI is rapidly reshaping how organizations search, reason over, and contextualize information, yet deploying these systems on private, confidential, and sensitive data introduces distinct architectural, performance, and governance challenges. This paper presents a practical roadmap for the 2026 and 2027 enterprise generative AI stack, with emphasis on agentic AI architectures built on retrieval-augmented generation (RAG), vector embeddings, and secure model deployment. Core concepts including encoding, similarity search, and vector databases are introduced to explain how knowledge is stored, retrieved, and reused across AI agents, alongside API-driven patterns that enable tools for search, reasoning, and contextualization. The discussion contrasts proprietary large language models (LLMs) and open-source small language models (SLMs), highlighting trade-offs in output quality, performance, and deployment, while examining how ecosystems such as Hugging Face support localized inference and domain-specific adaptation. Implementation considerations are presented primarily in Python, with extensions to R and SAS® Viya®, focusing on secure, reproducible analytics workflows. This paper concludes with prompt engineering strategies and ethical considerations for responsible generative AI use on sensitive enterprise data in the life sciences, pharmaceutical, and healthcare domains.

Developers, data scientists, and analysts are increasingly leveraging open-source tools and libraries to integrate with and enhance their existing data engineering and analytical workflows. One of these integrations, built into SAS® Viya®, is its pre-configured Python runtime integration, PROC PYTHON, that gives SAS programmers access to Python's open-source data science libraries for processing, visualizing, and analyzing data alongside SAS procedures. This presentation demonstrates how to access and use Python libraries in Viya runtimes; understand data-handling in Python and SAS; build Python scripts with reusable methods that import, process, and analyze data; and execute those Python methods to export Pandas DataFrames as SAS datasets.