Aparna Dhinakaran

Aparna Dhinakaran is the cofounder and chief product officer of Arize AI, where she builds tools that reveal how artificial-intelligence systems behave in production and why they fail. Her work spans AI observability, application-specific evaluation, and software agents that learn from their mistakes.

From production machine learning to Arize

Dhinakaran studied electrical engineering and computer science at the University of California, Berkeley, conducted research with Berkeley AI Research, and later entered Cornell University’s computer-vision doctoral program before taking a leave of absence. She worked at Apple and TubeMogul, then helped build machine-learning infrastructure at Uber, including its Michelangelo platform.

She subsequently led Monitor ML, a Y Combinator-backed machine-learning monitoring startup. When Arize acquired its team in 2020, she became the company’s cofounder and chief product officer. In her writing on opaque automated decisions, she connected model transparency to accountability in lending, insurance, hiring, and fraud detection.

As predictive models gave way to language models and agents, Arize expanded into tracing, evaluation, and experimentation. Its organization-owned open-source project Arize Phoenix helps developers inspect application behavior, assemble evaluation datasets, and test improvements. Dhinakaran and chief executive Jason Lopatecki announced the company’s $70 million Series C in 2025.

How she approaches reliable AI

• Evaluate the application, not the leaderboard. A highly ranked model can still route a shopping question to the wrong function or pass incorrect arguments to the right tool. Dhinakaran evaluates routing, retrieval, tool selection, and parameter extraction separately, using production traces to create datasets and experiments.

• Make evaluations explain their judgments. Binary labels and numerical scores rarely tell engineers what to fix; her experiments also found that substantially different document-error rates could receive similarly extreme grades. Explanations distinguish misunderstood intent, irrelevant retrieval, incorrect tool arguments, and other actionable failure modes.

• Inspect the entire agent trajectory. Reliable agents must select appropriate tools, execute them in a sensible order, retain conversational context, and finish without unnecessary loops. Her agent-evaluation framework also addresses voice-specific challenges, including transcription accuracy, speech quality, and conversational continuity. She argues that evaluation prompts require their own improvement loop because faulty judges can misclassify successful interactions.

• Use prompt learning to preserve useful feedback. Working with Claude Code and Cline, Dhinakaran’s team combined software-engineering tasks, unit tests, and explanatory model judgments to generate persistent CLAUDE.md instructions or Cline rules. This prompt-learning approach improves agent behavior without changing model weights.

• Deploy adaptive agent judges. Fixed rubrics struggle when agents generate interfaces dynamically, pursue long-horizon tasks, or repeatedly revisit tools. Arize’s Signal investigates production traces for inefficient trajectories and unexpected failure patterns, complementing deterministic tests and conventional model-based judges. Her work on self-improving agent systems links investigation, experimentation, evaluation, and repair.

These ideas depend on dependable infrastructure: her OpenTelemetry analysis examines immutable spans, delayed evaluation results, dropped attributes, and streamed responses. She has also credited Phoenix’s open-source lead and small engineering team, emphasizing that exceptional engineers build systems whose impact compounds.