Dat Ngo

Dat Ngo is Head of Agent Strategy at Arize AI as of September 2026. His public technical work explains how teams can change AI behavior safely: giving prompts their own operational lifecycle when needed, tracing outputs to the configuration that produced them, and measuring retrieval changes through repeatable experiments.[3][5][6]

In his April 2026 article “Prompt templates as configs, not code”, Ngo identifies a practical boundary: “A prompt template becomes config when it needs a lifecycle separate from the application deploy.” Frequent instruction changes, model swaps, or experiments can otherwise require rebuilding and deploying an application whose logic has not changed. He recommends keeping prompts in code when they change rarely, belong to a small stable workflow, or remain tightly coupled to application logic.[5]

Ngo’s proposed lifecycle is draft, evaluate, canary, promote, and rollback. He recommends shipping a known-good local default, allowing remote overrides, caching in the serving path, and attaching version metadata to every run. This gives teams a way to test behavior on a small traffic cohort and reverse a bad change without a full application release, while retaining a fallback when remote configuration is unavailable.[5]

The unit he proposes managing is broader than prompt text: it includes required inputs, model and inference settings, tool and retrieval assumptions, output expectations, and rollout metadata. Each run should be traceable to the configuration that produced it. He also draws a firm boundary around software guarantees: “Business logic, authorization rules, and application invariants should not be delegated to prompt config.”[5]

In March 2025, Ngo and Alejandro Cantarero of DataStax coauthored a RAG experimentation tutorial. They use SQuAD question-and-answer references to evaluate a Langflow chatbot backed by Astra DB and NVIDIA services, logging its answers and correctness scores in Arize Phoenix. Their evaluator uses an LLM judge distinct from the model generating the chatbot’s responses.[6]

The tutorial then inserts NVIDIA’s nvidia/rerank-qa-mistral-4b reranker between vector search and the prompt, expands the retrieved candidate set, and asks readers to repeat the experiment. The example teaches a concrete comparison workflow rather than reporting a quantified accuracy gain. Ngo also hosted the Retrieval+Search track at the 2025 AI Engineer World’s Fair.[6][1]

Before his Arize work, Ngo held data-science leadership roles at alliantgroup, PointPredictive, and Wood Mackenzie. He earned a master’s in Applied Statistics from Texas A&M University. His earlier Wood Mackenzie work provides context for his experience with production data pipelines.[2][7]

A presentation by Ngo and Yan Fung, updated September 25, 2019, describes a Kafka-based oil-and-gas data platform feeding Wood Mackenzie’s LENS. Their shared account covers scraping disparate sources, cleaning and validating data, deriving well attributes, and transforming inputs into individual well records. Among their lessons was to “know your data types/stores early (zero vs null, schema)”; a move from Google Protocol Buffers to Avro appeared as a possible future improvement.[7]