Aparna Dhinakaran

Where others compete on model scale, Aparna Dhinakaran has concentrated on what happens after deployment: tracing behaviour, evaluating failure and turning production evidence into better agent systems. Her work has helped make AI reliability an engineering discipline rather than a launch-day promise.

AI’s decisive battles are increasingly fought after a model ships: in traces, failure taxonomies, evaluator design and the difficult work of learning from production. Aparna Dhinakaran has made that operational layer her field. As co-founder and chief product officer of Arize AI, she has helped turn observability from a specialist discipline for predictive models into a working architecture for generative applications and agents. In 2025 she co-authored Prompt Learning, an approach that feeds natural-language critiques from evaluations and annotations back into system instructions. The company’s published tests reported a ten-percentage-point improvement on a selected Big-Bench Hard run, while keeping edits readable and reviewable rather than burying them in model weights. The same product year brought session- and trajectory-level evaluations, prompt optimisation, real-time tracing and 91 platform releases; the company reported 3.5 million monthly Python downloads for its open-source tooling and more than 567 billion ingested spans. By 2026, its open telemetry work was connecting production observability with framework-agnostic agent evaluation and controls. Dhinakaran’s contribution is not another model. It is the evidence layer that lets teams see, test and improve the systems models become.