Preetika Bhateja is a product manager and former Google Cloud strategic cloud engineer specializing in production-grade agent evaluation. Her work on generative advertising confronts a consequential reliability problem: ensuring AI-generated creative is accurate, brand-safe, and preserves legally required disclosures.
In 2021, she coauthored Google Cloud guidance on serverless data-pipeline orchestration, integrating Workflows, Dataflow, Cloud Storage, BigQuery, and Cloud Functions into recoverable, observable processing pipelines. In 2022, she coauthored a guide to automating BigQuery dataset snapshots for backup and recovery.
At AI Engineer World’s Fair 2026, Bhateja was identified as a Google product manager affiliated with YouTube Ads. Her joint conference session with Daniel Bump examined how evaluation systems mature alongside advertising agents.
Her approach to reliable advertising agents
- Human-calibrated evaluation rubrics: Establish agreement about successful outputs, document edge cases, and require explanations that distinguish product ambiguity from genuine agent failures.
- LLM-judge calibration: Compare automated judgments against expert-reviewed examples and track disagreement instead of trusting unexplained pass-or-fail scores.
- Agent trace analysis: Inspect intermediate decisions for failures invisible to aggregate metrics. Bhateja described an agent that detected a mandatory advertising disclaimer and then removed it despite explicit instructions to preserve it.
- Multidimensional creative evaluation: Score accuracy, brand safety, and other requirements separately; refresh test cases with production data and define which regressions must prevent launch.
