Ankur Duggal

Ankur Duggal is a solutions architect at Arize AI, helping enterprise teams investigate and improve AI agents and applications. His work connects agent observability with evaluation: understanding how an application reached an answer, locating the step that failed, and testing whether a change fixes the underlying problem.

Before Arize, Ankur designed and deployed voice agent platforms supporting hundreds of concurrent calls. He holds a master’s degree in data science from Indiana University. That earlier work involved building systems that handle many simultaneous conversations; his work at Arize centers on diagnosing and improving the behavior of enterprise AI applications.

Debugging beyond the final answer

Ankur’s approach starts with the gap between a plausible response and a dependable application. An answer can read convincingly while concealing fabricated information, repeated searches, or a tool failure that the agent silently worked around. Runtime traces expose the supervisor, specialist agents, model calls, and tools behind the response, giving developers a way to locate the failure rather than judge only the finished prose.

A concrete example from his work with a financial agent concerns stock queries outside the United States. Changing the country context was insufficient until the underlying tool was also updated. The distinction matters to his debugging approach: an agent’s instructions and the capabilities of its tools both shape what it can accomplish. A prompt edit cannot by itself supply a missing tool capability.

Turning failures into useful tests

Ankur uses Arize’s CLI and agent skills to bring traces and evaluation results into a coding agent’s workspace. That gives the coding agent application behavior to investigate alongside the code. Dataset, evaluator, and experiment workflows then make proposed changes testable against the failures that prompted them.

Two practical priorities shape this work:

• Application-specific evaluations. A fluent summary still needs checks against what it was supposed to report. Ankur recommends evaluations tailored to the application’s failures so that later changes in models or traffic do not quietly reintroduce them.
• Human review of proposed changes. Coding agents can spend their effort on linting while leaving the original behavioral defect unresolved. Ankur uses failed evaluations and traces to focus the investigation, while retaining human review of the resulting edits and tests.