David Smith is a statistician, open-source community builder, and AI engineering advocate whose work spans the development of the R ecosystem and production-ready generative AI. He co-authored An Introduction to R, helped establish the R Consortium, and later taught developers to build and evaluate AI applications grounded in business data.
From R to applied AI
Smith studied statistics at the University of Adelaide and Lancaster University. His work with Bill Venables at Adelaide helped form the R Project’s official introductory manual, and he contributed to the early development of Emacs Speaks Statistics.
He directed product management for S-PLUS at Insightful before becoming chief community officer at Revolution Analytics, where he wrote extensively about R and data science for the Revolutions blog. In an essay on open-source data science, he argued that shared software, inexpensive infrastructure, and collaborative development made large-scale analytics economically viable.
After Microsoft acquired Revolution Analytics in 2015, Smith moved into Microsoft’s developer ecosystem. He helped launch the R Consortium that year and subsequently served on its board as treasurer. His later work at Microsoft and GitHub included leading AI and machine-learning advocacy and serving as a principal AI advocate. His personal GitHub projects include Copilot for R and prompt-engineering workshops. He has since announced his departure from Microsoft and GitHub; no subsequent employer is independently established.
- Hybrid retrieval for production applications. Smith recommends combining keyword and vector search, arguing that the combination can retrieve relevant information more effectively than semantic similarity alone. His retail-copilot workshop, co-led with Cedric Vidal, combined product search with customer purchase histories to generate personalized answers.
- Groundedness as an engineering requirement. Smith evaluates model responses for coherence, fluency, relevance, and factual grounding. His example of a chatbot inventing a nonexistent toothbrush shows why polished language cannot substitute for checking answers against actual source data. The Contoso Chat sample demonstrates the associated retrieval, deployment, and evaluation workflow.
- Continuous evaluation and prompt-injection awareness. Smith treats production AI as a cycle of prompt refinement, retrieval tuning, adversarial testing, monitoring, and reevaluation as models or data change. He highlights prompt injection as a particular concern when user input influences generated database queries, alongside access controls and customer-data boundaries.
