There is also a data engineering part to this. For example, developing a data pipeline for ground truth. How would you go about this?

To build a ground‑truth pipeline I would start with a clear definition of the target data and the validation rules. First, ingest raw data from the source system into a staging area (e.g., S3 or a data lake). Next, apply schema validation to ensure the data matches the expected format and types. Then, perform data‑quality checks such as null‑value detection, range checks, and consistency checks across related fields. For labeling or annotation, I would use a semi‑automated approach: pre‑label with a model, then have human reviewers correct or confirm the labels. After labeling, store the validated data in a versioned table (e.g., Delta Lake or Parquet) and publish it to downstream services. Finally, set up monitoring and alerting (e.g., CloudWatch or Grafana) to detect drift or new data quality issues. This end‑to‑end pipeline ensures that the ground truth is accurate, reproducible, and available for training or evaluation.