# From notebook to production

A pack of 8 sessions from the MLOps Community YouTube channel, in the order to watch them. 7h 08m of video.
Page: https://mlopstalks.com/packs/from-notebook-to-production

The model works when you run the notebook. Someone else still needs your help to find the input files, execute cells in the right order, and turn the result into something the business can use again. The first production milestone is a workflow that survives that separation from its author. Start by identifying what comes after training, then make hidden state and assumptions explicit. Extract reviewable code, separate configuration from processing, and try running the project elsewhere. An audio example shows why the preprocessing must travel with the model. Virgin Media adds a production-like test environment; Regions Bank finishes with training, scoring and monitoring as repeatable applications. The older demonstrations supply design examples, not a requirement to adopt their particular tool stacks.

## This pack is for you if

- A colleague cannot run your successful notebook without asking you which cells and files to use.
- You have a trained model, but no repeatable way to turn fresh inputs into usable predictions.
- Your first deployment needs a maintainable project rather than another manual export.

## The talks, in order

### 1. End to End MLOps Basics

Raviraja Ganta, Enterpret | 57:59 | MLOps Meetup
Video: https://www.youtube.com/watch?v=B1t_Vb2MkRw
Summary: https://mlopstalks.com/talks/end-to-end-mlops-basics.md

Why first: A successful training run is one stage of a longer job. Ganta separates development, packaging, serving and maintenance so you can name what your notebook has yet to become. His 2021 stack illustrates those responsibilities; start with the missing work before choosing tools for it.

### 2. I Don't Like Jupyter Notebooks

Joel Grus, Capital Group | 55:41 | MLOps Coffee Sessions
Video: https://www.youtube.com/watch?v=3Fa6uzHxTkQ
Summary: https://mlopstalks.com/talks/i-dont-like-jupyter-notebooks.md

Why here: Hidden cell state makes even the starting result difficult to explain. Grus describes an experienced engineer losing track of variable values, then shows how his team tested the same training image locally before sending it to the cloud. That gives the transition a concrete test: can it run outside your interactive session?

### 3. Challenges Operationalizing ML (And Some Solutions)

Nathan Ryan Frank, WW Grainger | 52:28 | MLOps Podcast
Video: https://www.youtube.com/watch?v=ezWdXzhwH8A
Summary: https://mlopstalks.com/talks/challenges-operationalizing-ml-and-some-solutions.md

Why here: The next move is a small extraction into shareable project code. Frank recommends packages, modules and tests while the reasoning behind the experiment is still fresh. This supplies the practical bridge between recognizing hidden state and adopting a more explicit project structure.

### 4. Building 12-Factor Data Apps with Kedro

Ivan Danov, QuantumBlack | 1:22:18 | MLOps Meetup
Video: https://www.youtube.com/watch?v=j-qPZkdJREw
Summary: https://mlopstalks.com/talks/building-12-factor-data-apps-with-kedro.md

Why here: Named inputs and outputs make that project easier to move. Danov's Kedro example keeps processing in Python functions and data locations in configuration. The distinction lets you change a storage location without rewriting the calculation, and prepares you to inspect the actual files in the next demonstration.

### 5. So Fresh and So Data Clean

Tommy Dang, Mage | 49:58 | MLOps Meetup
Video: https://www.youtube.com/watch?v=6Iyt9Wip3C4
Summary: https://mlopstalks.com/talks/so-fresh-and-so-data-clean.md

Why here: Here is the transition as a working demonstration. Dang commits focused Python files and dependency metadata, then runs the project on another server. His 2022 Mage demo keeps interactive exploration while making the kept work reviewable; the lesson is what reaches the repository and runner, not the historical installation.

### 6. Bringing Audio ML Models into Production

Valerio Velardo, Utopia Music | 50:42 | MLOps Coffee Sessions
Video: https://www.youtube.com/watch?v=fGKZljg_SQI
Summary: https://mlopstalks.com/talks/bringing-audio-ml-models-into-production.md

Why here: A portable project still needs the right preprocessing. Audio makes the omission obvious: wave files must become suitable representations before many models can use them. Velardo argues for shared transformations alongside training, evaluation and deployment, so the production artifact includes the work that made the experiment meaningful.

### 7. From Notebooks to Production FASTER

Shahd Alghrsi, Virgin Media | 13:43 | MLOps Community
Video: https://www.youtube.com/watch?v=4JNvJwnXmcA
Summary: https://mlopstalks.com/talks/from-notebooks-to-production-faster.md

Why here: Now test the assembled project somewhere that resembles production. Alghrsi describes Virgin Media's templates, container images and separate experiment, UAT and production environments. This adds a controlled place to discover environment differences before the repeatable applications in the final talk become operational.

### 8. Operationalizing Machine Learning at a Large Financial Institution

Daniel Stahl, Regions Bank | 1:04:58 | MLOps Meetup
Video: https://www.youtube.com/watch?v=vrvagiFVzI4
Summary: https://mlopstalks.com/talks/operationalizing-machine-learning-at-a-large-financial-institution.md

Why last: Regions Bank runs the training process again in production rather than relying on the model left over from exploration. Stahl's training, scoring and monitoring pipelines show what the notebook has become: versioned code that can produce, use and check a model repeatedly. His 2021 account closes the route with the repeatable work needed to keep using the experiment's result.

## Editor's note

Frank wants notebook work to become tested, shareable code; Stahl wants the full training process to run again in production. ZenML expresses workflows as Python steps and lets the same pipeline run on different infrastructure through configuration. Each run records its steps, inputs, outputs and code version, so the resulting model can be traced to the work that produced it.

Written by the MLOps Talks editors (the ZenML team), not by any of the speakers.