Latest in AI
NVIDIA Cosmos Data Factory Setup: Check the Cluster Before the Two Prompts
Cosmos can orchestrate synthetic-data workflows, but read-only readiness checks, deployment permissions and a realistic cost boundary come first.
NVIDIA's November 18 Cosmos Data Factory walkthrough presents five synthetic-data blueprints and six reusable agent skills. They cover tasks such as video augmentation, rare-event generation and inspection defects. The attractive part is coordinating the pipeline through an agent; the part to inspect carefully is what that agent will change in your infrastructure.
I would separate the first trial into an audit, a reviewed deployment and a small generation job. Combining those into one broad instruction makes it harder to notice which step needs credentials, creates a bill or interrupts existing work.
Start with the workload, not the headline
The orchestration repository uses Airflow on Kubernetes and can deploy inference services and batch tasks. Some workflows also support OSMO, where they use external model endpoints. The platform choice affects what must already exist and what the workflow can manage itself.
Choose one concrete data gap. For example, you may need additional lighting conditions for an existing inspection scene, rather than an entire new dataset. Define the intended change, the properties that must remain stable and where accepted outputs should go.
That brief becomes your evaluation contract. A generated image looking plausible is not enough if it silently changes the component whose defect you are trying to detect.
Run an audit without authorizing repairs
The preflight skill starts with a read-only audit and separates controller installation readiness from workflow readiness. UNVERIFIED means evidence is missing. PARTIAL can mean one workflow is ready while another is blocked.
An original starting instruction would be:
Audit this checkout and the specified cluster without changing either. Report the selected cluster, installation readiness and readiness for my chosen workflow. List missing prerequisites and proposed changes separately. Do not install packages, alter storage, deploy services or start generation.
This is a proposed prompt, not a tested run. Give the agent the intended cluster configuration explicitly, and review the report before approving any remediation. Do not paste credentials into its summary or a shared issue.
A passing audit is useful evidence about a particular configuration at a particular time. Save that configuration with the report so a teammate can understand what was checked.
Review the destructive installation boundary
The setup guide warns that make install sdg-controller first uninstalls existing resources and deletes workloads and services across the target namespace. Persistent volume claims remain. This is not a harmless refresh while other people are running jobs there.
The same guide says make setup rewrites its managed secrets files rather than merging arbitrary custom values. Setup itself does not deploy Airflow; building and installing are separate steps. It also documents an Airflow port-forward reachable beyond localhost and default credentials that must be changed before production use.
I would use a dedicated namespace for the trial and schedule installation when it contains no active work. Review the actual target before executing the command, back up configuration you need to preserve, and have a named owner for cleanup. Never treat an agent's successful command exit as evidence that unrelated workloads were unaffected.
Check the complete cost boundary
NVIDIA's walkthrough bases its compute examples on an eight-GPU server at about $40 per hour, or $960 per day. It explicitly excludes costs such as setup, source preparation, storage, agent use and human review.
For an arithmetic example, a $960 compute day that yields 100 accepted assets costs $9.60 per accepted asset before those other expenses. If only 50 outputs meet your requirements, that becomes $19.20. These are hypothetical acceptance counts, not measured Cosmos results.
Set a small initial output limit and include rejected generations in the cost record. Otherwise, cheap generation can hide an expensive review process. Confirm the selected workflow's actual resource requirements rather than assuming every blueprint needs the same allocation.
Validate against the real task
Keep source assets, generated variants and evaluation records connected. Review whether the requested change occurred, whether labels describe the result correctly and whether important scene properties survived.
Then assess usefulness on held-out real data, not just generated examples from the same pipeline. A synthetic-data workflow earns its place when it helps the downstream model handle the missing case. The number of prompts needed to launch it is a convenience, not that evidence.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help AI teams explain complex deployment workflows and product tradeoffs clearly enough for developers to act on them.