Taylor Jordan Smith is a principal developer advocate at Red Hat who helps enterprises build reliable, governable generative AI systems. Her work centers on production LLM evaluation, self-hosted inference, AI agents, and open-source infrastructure.
With a background in software development, technical partnerships, Linux, and Kubernetes, Smith approaches AI deployment as an infrastructure problem as much as a model-quality problem. Her 2025 AI Engineer World’s Fair workshop demonstrated practical evaluation using GuideLLM, vLLM, the Language Model Evaluation Harness, MMLU-Pro, and Promptfoo. By 2026, her focus also encompassed agent reliability, distributed inference, and governed access to enterprise-hosted models.
- Evaluate the complete production system. Smith organizes evaluation in layers: throughput, latency, concurrency, accelerator utilization, and cost come first; factual accuracy, output formatting, safety, bias, and application-specific behavior follow. She recommends testing individual components before expanding into integration and end-to-end workflows, then incorporating evaluations into continuous delivery.
- Benchmark realistic workloads. Using GuideLLM and vLLM, she varies request rates and input-output token configurations to expose time to first token, inter-token latency, and tail performance. Useful benchmarks reflect the actual demands of retrieval systems, coding assistants, and customer-facing applications.
- Measure whether agents finish the job. Her approach to agent evaluation considers task completion, tool and Model Context Protocol correctness, recovery from failures, multistep reliability, latency, human review, tracing, and regression analysis.
- Treat private model access as shared infrastructure. Smith describes enterprise Models-as-a-Service built around governed internal APIs, OpenShift, distributed inference with vLLM and llm-d, NVIDIA infrastructure, observability, and GPU telemetry. Self-hosting offers organizations greater control over data boundaries, costs, latency, and governance while introducing operational responsibilities that hosted APIs avoid.
