FDE Times: DeepLearning.AI LLMOps course — keep the rerunnable pipeline, drop BigQuery when the client isn’t on GCP
A 5 November FDE Times review of Erwin Huizenga’s ~1h31m DeepLearning.AI + Google Cloud LLMOps course says the FDE takeaway is a supervised fine-tuning pipeline that versions data and models and scores safety per category—not Vertex or BigQuery. On a non-GCP client, swap the data store, keep Kubeflow Pipelines (or their orchestrator), and put per-category thresholds in the acceptance criteria.
DeepLearning.AI’s LLMOps short course (with Google Cloud, taught by Erwin Huizenga) runs about 1 hour 31 minutes and centres on adapting an open-source supervised fine-tuning pipeline on BigQuery, Kubeflow Pipelines and Google Cloud. FDE Times (5 November 2026) argues the client-facing value is not the GCP toolset: many deployments sit on AWS, Azure or a private Kubernetes cluster.
Three habits travel. First, a compiled pipeline the customer’s team can rerun beats a notebook only you can run. Second, version both the dataset and the model on every tune so a worse v3 can be compared to v2 instead of guessed about. Third, output a safety score per harm category rather than a single safe/unsafe bit—then put each threshold in the acceptance criteria the client signs.
Why it matters for FDEs: when the client isn’t on GCP, replace BigQuery with their store, keep Kubeflow Pipelines (or map steps into the orchestrator they already run), and treat versioning plus per-category safety as portable practice. A one-page rewrite that an on-site engineer can rerun without you is the real exit criteria.
Source: fdetimes.net