Workflow · Production

End-to-end ML pipelines with Kubeflow, Docker, uv, and Vertex AI for local and cloud LLM experimentation

The problem

Data scientists face the challenge of bridging local development with scalable cloud deployment — the 'works on my laptop' problem — while LLMs add costs and development time if pipelines are not managed correctly, and slow remote feedback loops make iteration painful.

Workflow diagram · grounded in source
1
Pipeline triggered
trigger
“The end goal is to be able to trigger our pipeline with a single command”
2
Get data
integration
“# Step 1: Get data data_task = get_data(data_source=job_params.data_source)”
3
Call LLM
ai_action
“# Step 2: Call LLM with the data llm_task = call_llm(model_name=job_params.model_name, prompt=data_task.output)”
4
Evaluate results
validation
“# Step 3: Evaluate results evaluate_results(results=llm_task.output)”
5
Compare experiments
feedback_loop
“we track all input parameters, save all intermediate artefacts, and log all metrics, greatly empowering our experimentation capability”
Reported outcome

The pipeline setup enables local execution with feedback within seconds, remote builds completing in under one minute due to caching, and experiment tracking to compare LLM runs across parameters.

Reported metrics
Remote build time with cachingunder one minute
Local container rebuild timeless than ~10 seconds
Reported stack
Kubeflow PipelinesDocker DesktopuvVertex AIBuildKittypermypyGitHub CI/CD
Source
https://mlops.community/blog/ml-pipelines-in-the-age-of-llms-from-local-containers-to-cloud-experiments
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The pipeline setup enables local execution with feedback within seconds, remote builds completing in under one minute due to caching, and experiment tracking to compare LLM runs across parameters.

What tools did this team use?

Kubeflow Pipelines, Docker Desktop, uv, Vertex AI, BuildKit, typer, mypy, GitHub CI/CD.

What results were reported?

Remote build time with caching: under one minute; Local container rebuild time: less than ~10 seconds (source-reported, not independently verified).

How is this workflow AI workflow structured?

Pipeline triggered → Get data → Call LLM → Evaluate results → Compare experiments.