From notebook to deployment with Vertex AI pipelines

This talk was held at the University of Pretoria, at the Google Developer Group (GDG) Pretoria for DevFest 2025, titled "From notebook to deployment with Vertex AI pipelines". The talk discussed using Paper Mill to productionise models and deploy them via Vertex AI as a custom pipeline.

Transcript

1. From "Senior Engineer", with the help of the Vertex AI 2025 Thamu Mnyulwa Machine Learning Engineer From notebook to deployment with Vertex AI pipelines Thamu Mnyulwa Machine Learning Engineer 2025
2. Hi I’m Thamu
3. Hi I’m Thamu Student in Applied Math Machine Learning Engineer Data Scientist I’ve been lucky enough to learn and work using GCP to automate models over the past 4 years This talk is in a personal capacity and does not reflect the views or opinions of my employer
4. Hi I’m Thamu Student in Applied Math Machine Learning Engineer Data Scientist Today’s talk covers how to use Vertex AI and Papermill to automate notebooks as Vertex AI jobs that you can run with your personal GCP account
5. 1. Who has worked with notebooks and deployed a model ?
6. 1. Who has worked with notebooks and deployed a model ? 2. Who has automated or scheduled that model to run?
7. Value in the real world comes when models are put out in the wild (in production) The goal of this talk is to introduce some principles of how to automate notebooks using Vertex AI and limited resources
8. Let’s walk through the journey of automating a notebook And use a small abstracted problem
9. The Journey • You build a model in a Jupyter notebook • It works beautifully locally • Now you need to run it daily on new data • How do you get from notebook to be automated?
10. Why Notebooks Are Powerful • Great for exploration and experimentation • Visual feedback and documentation • Familiar to data scientists • Easy to share and reproduce
11. Why Notebooks Are Powerful • Great for exploration and experimentation • Visual feedback and documentation • Familiar to data scientists • Easy to share and reproduce • Needs to run automatically, not interactively • Must handle different parameters and data • Requires proper error handling • Needs monitoring and logging But Production is Different
12. A Solution • Split notebooks into operational stages • Use Papermill to parameterize execution • Deploy on Vertex AI Custom Jobs • Automate with proper ML pipelines
13. Start off with a simple model in a single notebook 1⃣
14. Break it into individual sections/notebooks [fs_train, train, fs_score, score] 1⃣ 2⃣
15. Automate via Papermill from and a Bash script https://papermill.readthedocs.io/ 1⃣ 3⃣
18. 3⃣ 4⃣
19. 5⃣
20. 6⃣
21. Show Me the Code
22. Vertex AI pipeline (might be your go to solution in an enterprise) • Vertex AI is the AI/ML platform to run AI jobs in Google Cloud • You can define your notebook as ◦ a Kube Flow pipeline or ◦ TensorFlow Extended • Vertex AI Endpoints are the end result • Qoutas will not allow you to take this route for personal projects
23. • There comes a time in life when you need to look for an answer in an unconventional way • Consider containerising most applications or projects before deployment Now what can I do?
26. When do we use Vertex AI vs Cloud Run Is Vertex AI fit for purpose for your use case?
29. Vertex Pipeline Approach (many nodes)
30. Final solution includes more considerations
31. (keep) (it) (simple) When in doubt simplifying the intended architecture is a good principle Conclusion
32. Guiding principles and layers for implementation
33. Use Gemini to experimentation whilst learning When in doubt you can use Gemini to help you convert notebooks into pipelines • Veo 3 was released this week • Google Antigravity with Veo3 free* ◦ Google’s AI IDE • Gemini CLI has Veo3 free* * for now 21 November 2025
34. • Papermill lets you parameterize and automate notebooks. • After containerizing your model with Docker and pushing the image to Artifact Registry, you can run it as a Vertex AI custom job. • Since Vertex AI compute quotas can’t be increased in a region without business justification, custom jobs are a practical option for personal projects. • Gemini CLI and Antigravity (the IDE) can help you iterate Summary
35. Questions Thamu Mnyulwa https://x.com/thamu_mnyulwa
36. Thank you! Thamu Mnyulwa https://x.com/thamu_mnyulwa …also, if you want to chat more about GCP let’s connect
37. List resources —--------------------- • https://home.mlops.community/public/videos/integrating-knowledge-graphs-an d-vector-rag-for-efficient-information-extraction * for now 21 November 2025