Aparna Dhinakaran
A speaker and practitioner in the AI engineering field, active in the AI engineer community with a focus on AI evaluation and engineering practices.
Timeline (last 90 days)
Dhinakaran argues that AI engineers should spend more time writing evals and need to "evaluate the evaluator": build golden-standard datasets, validate eval results against human outputs, and only release a new evaluator after ensuring the judge aligns with human judgment.
Dhinakaran believes the core capability of AI engineers is newly defined as "defining what good looks like"; as agents take on more and more execution work, the value of human engineers concentrates on establishing and calibrating judgment criteria.
Dhinakaran suggests tracking not only the agents being built but also the coding agents that build them—the call chains of Claude Code, Cursor, and Codex should all be recorded.
Dhinakaran acknowledges that an obvious gap exists between lab environments and real enterprises, where data is far messier than the controlled, measurable data in labs.
All Facts (4)
Dhinakaran argues that AI engineers should spend more time writing evals and need to "evaluate the evaluator": build golden-standard datasets, validate eval results against human outputs, and only release a new evaluator after ensuring the judge aligns with human judgment.
50%UnverifiedDhinakaran believes the core capability of AI engineers is newly defined as "defining what good looks like"; as agents take on more and more execution work, the value of human engineers concentrates on establishing and calibrating judgment criteria.
50%UnverifiedDhinakaran suggests tracking not only the agents being built but also the coding agents that build them—the call chains of Claude Code, Cursor, and Codex should all be recorded.
50%UnverifiedDhinakaran acknowledges that an obvious gap exists between lab environments and real enterprises, where data is far messier than the controlled, measurable data in labs.
50%