From AI Canvas to MLOps Stack Canvas: Are They Essential?

Author:Murphy | View: 29135 | Time: 2025-03-22 19:56:13

Hey! I have been using the AI and ML canvases for both personal and professional projects for some time now, as it helps me keep my projects and ideas well-organized. Lately, I have been thinking about creating a canvas that is MLOps-oriented and while doing some research on the topic, I found a ready-made canvas called "MLOps stack canvas". In this article, we will explore together how to use the MLOps stack canvas. At the end, I'd love to hear your thoughts on the canvas; and let me know if you have used it before or if you plan to adopt it in the future!

Not a Medium member? No worries! Continue reading with this friend link.

Table of contents:

· 1. Introduction · 2. AI canvas · 3. ML canvas · 4. MLOps stack canvas · 5. So AI, ML or MLOps canvas? · 6. Conclusion

My MLOps tutorials:

[I will be updating this list as I publish articles on the subject]

1. Introduction

In general, a canvas is a visual framework that helps individuals and teams to map and analyze various aspects of a given project (not necessarily a computer science project) in a structured manner. In the context of Machine Learning, it is a tool to outline, design and communicate various components of the ML project: data preparation, data preprocessing, training, evaluation, etc.

In MLOps, the first step to execute before diving into code is to thoroughly understand the problem that we aim to solve. This latter involves defining the inputs, the outputs, the ML model, etc. Since there are many components to define, it is important to outline them clearly for easy reference. The AI (Artificial Intelligence) canvas or the ML (Machine Learning) canvas can serve as templates for organizing these project components. Recently, as MLOps has gained industry attention, the MLOps Stack Canvas was introduced to further improve this process.

In this article, we will explore the AI, the ML and the Mlops Stack canvas respectively and illustrate their application with examples. After that and before concluding this article, we will discuss their importance, whether their use is essential, and whether relying on one or some of them is sufficient.

I usually use the same example throughout this series of articles, but today we will explore a different one. So by the end of this series, you will have two distinct examples. Sounds good, right? The example we will be treating today is: Comment Moderation in Social Media (CM-SM). This involves monitoring, reviewing and managing user comments on social media platforms. The goal is to implement a machine learning model to automatically review and manage the comments to ensure that discussions remain respectful and compliant with the community guidelines and legal regulations.

2. AI canvas

AI canvas proposed in the book "Prediction Machines. The Simple Economics of Artificial Intelligence" [1] with three additional fields inspired from the business model canvas and the fields numbered in the order they should be filled as an additional update.

The AI canvas was first proposed in "Prediction Machines. The Simple Economics of Artificial Intelligence" [1], a book for explaining the implication of AI in the business world including strategies and decisions. AI canvas is a tool developed to help decompose the tasks of AI-based problems' solutions. It is one of the components that I presented in my previous article as a template for organizing the MLOps workflow.

In the book, the AI canvas was defined as :

.. an aid for contemplating, building, and assessing AI tools. It provides discipline in identifying each component of a task's decision. It forces clarity in describing each component [1].

Simply put, I see the AI canvas as a visual tool that helps us think about the workflow from creation to evaluation of AI systems. It's a structured and a clear way for breaking down the different steps of developing the AI system. It generally provides a high-level structure for ML/AI implementation.

Developing the AI canvas involves identifying the following components by answering their respective questions:

1. Action: What are you trying to do? Here, we identify the action to take according to the predictions. Example: The AI system automatically flags, hides or removes inappropriate comments.
2. Prediction: What do you need to know to make the decision? Here, we specify what we are trying to predict. Example: The AI model predicts whether a user comment is inappropriate or appropriate for the social media platform.
3. Judgment: How do you value different outcomes and errors? After the prediction results, we study, compare and evaluate the cost of responding in different situations. Example: The AI model prioritizes reducing the false negatives, inappropriate comments detected as normal, while monitoring the false positives, normal comments predicted as inappropriate to prevent disruption to genuine conversations.
4. Outcome: What are your metrics for task success? Example: Maintain a healthy and respectful online environment, increase user experience and reduce the need for manual flagging by moderators.
5. Input: What data do you need to run the predictive algorithm? Example: Comment text and comments visual content if included.
6. Training: What data do you need to run the predictive algorithm? Example: The system is trained using labeled user comments.
7. Feedback: How can you use the outcomes to improve the algorithm?Example: Gather feedback from users when they report inappropriate comments and feedback from moderators by reviewing the comments and predictions.

By addressing the previous questions, the resulting AI canvas appears as follows:

The AI canvas for: Comment Moderation in Social Media.

Hmm, the AI canvas provides the design and the understanding of the AI system. However, it lacks the details about the implementation and deployment steps for the entire ML pipeline, don't you agree? Can these details be found in the ML canvas? Let's find out in the next section.

3. ML canvas

The ML canvas as proposed in [2], with the fields numbered in the order they should be filled as an additional update.

As the AI canvas, the ML canvas is a tool designed to help decompose AI-based problem-solving tasks, but with more specification of the entire ML system. It was proposed by Louis Dorard as a…

Design framework to connect a value proposition to a Machine Learning task, specify requirements and anticipate costs. Recommended by leading companies and taught at top universities throughout the world. [2]

In a simpler way, I describe the ML canvas as a visual detailed guide for managing the ML lifecycle, including feature engineering and model training and evaluation. The ML canvas fields can be organized into four distinct logical groups: the objective group (value proposition), the learn group (in blue: data sources, features, data collection, building models), the predict group (in purple: prediction task, offline evaluation, decisions, making predictions) and the evaluation group (Live monitoring). Developing the ML canvas involves identifying the following components by answering their respective questions:

Value proposition: What are the objectives of this system? and what is the workflow? Example: The objective of CM-SM is to maintain a healthy and respectable environment for users by predicting the new comment class and hide it or remove it if it is inappropriate.
Data sources: From where the data is collected? Example: The data is collected from publicly available datasets and from the platform comment data.
Prediction task: what is the type of the task? specify the input and the output. Example: The model predicts from the comment text and the description of the visual content whether it is appropriate, not appropriate or a spam (multi-class classification).
Features: What is the input representation at prediction time? Example: The text of the comment, which will be transformed into word embeddings, the visual content (images), if available, that will be represented through word embeddings based on its description.
Offline evaluation: How the system is evaluated once trained/deployed? Example: At prediction time, the word embeddings are generated from the comment text and the description of the visual content.
Decisions: what is the decision to take according to the prediction output? Example: Hide the comments predicted inappropriate with low probability, remove comments predicted inappropriate with high probability, hide comments predicted as spam.
Making predictions: How are the predictions made once deployed? Example: The predictions are made on new comments only (real-time) and on all comments on all posts of the platform members (batch).
Data collection: How the data is collected? and how it's updated? Example: The collected comments are labeled and updated by the reported and reviewed comments.
Building models: How many models are needed in production? and when are they updated? Example: A single model for the predictions and the model is updated when moderators or users report more than K (a fixed threshold) comment.
Live monitoring: How is the model monitored in production? What is its impact in production? Example: Track the system predictions on the new comments.

By addressing the previous questions, the resulting ML canvas appears as follows:

The ML canvas for: Comment Moderation in Social Media.

You can request the original canvas from here and they will also provide you with another practical example. For further reading you can also check out these Medium articles written by, [Louis Dorard](), the creator of the ML canvas: From Data to AI with the Machine Learning Canvas (Part I) and From Data to AI with the Machine Learning Canvas (Part II).

Hmm, indeed, the ML canvas provides more details about the implementation of the ML model than AI canvas, but you feel like something is missing, don't you? Yes, that's exactly it! Additional details are indeed lacking, such as the deployment and the monitoring of the ML model that are covered in the MLOps Stack canvas, as we will see shortly!

4. MLOps stack canvas

The MLOps Stack canvas is a tool designed to outline the architecture and the infrastructure stack of MLOps within the ML project. It was first proposed here [3] and was developed to be adaptable for any application across diverse industries. The MLOps canvas consists of four different category: general management (value propositions, MLOps dilemmas), data and code management (data analysis & experiment management, data sources & data versioning, feature store & workflows, foundations), model management (CI/CT/CD: ML pipeline orchestration, model registry and model versioning, model deployment, prediction serving; Model, data and application monitoring) and metadata management (metadata store).

Drawing up the ML canvas involves identifying the following components by answering their respective questions. Note that, unlike in the previous sections, I will only explain the fields in the following points without providing examples to keep brief. But, don't worry, you will find the complete example in the MLOps Stack Canvas below.

1. Value proposition: As in the ML canvas, value proportion describes the purpose of the ML system in regards to the end-user but in a more elaborated and technical way. Filling out this component involves answering these questions: What is the objective of the system? What is the problem we are solving? What is its importance? Who are the team members involved in this project? and who will own the models in production?
2. Data sources and data versioning: This field focuses on estimating the cost of the different steps related to raw data including: acquisition, storage, preprocessing and versioning. Filling out this component involves answering these questions: Is data versioning mandatory? From where the data is collected and where it's stored? Does it require manual labeling?
3. Data analysis and experiment management: This field focuses on the ML tools required for effective data analysis and managing experiments (data preparation). Filling out this component involves answering these questions: What is the programming language used for data analysis? What are the infrastructure requirements? What are the evaluation metrics? And what ML framework is used?
4. Feature store and workflows: This field focuses on whether engineering needs to be reproducible, the infrastructure requirements and whether to build or buy a feature store. This block comes after data preparation. It also involves the details about transforming data into feature vectors in both training and production.Filling out this component involves answering these questions: Is it mandatory? How are features computed? What are the requirements of the infrastructure for feature engineering?
5. Foundations: This field focuses on bringing the DevOps best practices to the MLOps stack including code maintenance, system performance monitoring; and deployment, testing and tracking automation. Filling out this component involves answering these questions: How is code maintained? What source version control is used? Is notebooks versioning needed? How is the system performance monitored? What CI/CD pipeline for the code is applied?
6. Continuous Integration, Training, Deployment: This field focuses on applying CI/CD to the ML model while introducing CT (Continuous Training) as a new practice in MLOps that extends beyond traditional DevOps workflows. Filling out this component involves answering these questions: How is the model retrained? When? Where? And what is the trigger for it?What is the ML pipeline?
7. Model registry and model versioning: This field focuses on how the ML model is maintained and managed in terms of registry, versioning and storage. Filling out this component involves answering these questions: Is it mandatory? Where the ML model is stored and tracked? How is it versioned? In short, this block involves the details about the ML model after training.
8. Model deployment: This field focuses on the different steps required for deploying the ML model including the delivery format, the target environment , the release policy and the deployment strategy. Filling out this component involves answering these questions: What is the delivery format? What is the target environment to serve predictions? What is the model release policy? What is the deployment strategy? In short, this block involves details about deploying the model.
9. Prediction serving: This field focuses on the specifications for serving the ML model in the production environment for making predictions. Filling out this component involves answering these questions: What is the serving mode? Is distributed model serving required?
10. Model, data and application monitoring: This field focuses on the monitoring step after deploying the ML system. It mainly includes collecting metrics for model performance decay and data skew detection, as well as alerting and triggering strategy. Filling out this component involves answering these questions: Is it mandatory? What are the metrics that need to be collected? How to detect the model performance decay? How to detect the data skew? What are the operational aspects that are needed to be monitoring?What is the alerting strategy?
11. Metadata store: This field focuses on metadata and documentation management. Filling out this component involves answering these questions: What kind of metadata needed to be collected? What is the documentation strategy?
MLOps Dilemmas: This field focuses on the requirements for developing the ML project to provide good estimation of associated costs. Filling out this component involves answering these questions: What MLOps tools and platforms should be used? What type of skills are needed?

By addressing the previous questions, the resulting MLOps Stack canvas appears as follows:

The MLOps Stack canvas for: Comment Moderation in Social Media.

Hmm, absolutely! The MLOps Stack Canvas provides a deeper insight into the workflows and infrastructure necessary for a successful ML project. Now, the question is: Are these canvases mandatory? What is the relationship between them? Is one of them sufficient on its own? If so, which one is it? Without delay, let's discover it now!

5. So AI, ML or MLOps Stack canvas?

Aaaand now, we have reached the most crucial part of this article! Since the title of this section is a question, I'd like to respond with a series of question-and-answer pairs. Yes! I'm answering a question with another! It's becoming a hobby of mine, I guess!

What Values do they offer? The AI, ML and MLOps stack canvases offer various values that enhance the development and deployment of the ML systems. They bring organization, clarity, flexibility, focus and efficiency, and improve communication and collaboration.

Are they permanent? The AI, ML and MLOps stack canvases are designed to be neutral to application and industry, so they are rarely modified by adding or ignoring certain fields. However, their content isn't definitive and can be updated by the time, thus the importance of the field "version". In my experience, we rarely need to push the adaptation to adding/removing fields but we sometimes ignore some questions or address them in newer versions.

There are some similar fields between them; is it a repetition? Let's take the "Value Proposition" field that exists in all three canvases (in the AI canvas, this is called "Action"). If we check this field only visually, we see directly that the content is different from one canvas to another. In addition, if we take a deep understanding of the content, we conclude that it describes the purpose of the ML system in relation to the business in the AI canvas, in relation to the end user in the ML canvas and in relation to the ML system and techniques in the MLOps Stack canvas. Therefore, this is not a repetition.

What is the relationship between these canvases? Is one of them sufficient on its own? If so, which one is it? They are rather complementary: each canvas focuses on a specific aspect of the ML system and is intended for particular professional roles as shown in the figure below.

Are they mandatory? This is indeed a crucial question. Actually, it depends on the size and complexity of the project: some small projects only need AI, some others also need ML canvas. For example, in one of my previous articles, I only used the AI canvas to describe the project instead of writing lines of text. However, as the project gets bigger, other canvases are also needed. Therefore, for medium and large projects, I think the three canvases are really useful tools: they define the maturity of the ML project.

The degree of abstraction of the AI, ML and MLOps Stack canvas

6. Conclusion

Here comes the end of this article! In this article, we introduced three different visual tools of MLOps: the AI canvas, the ML canvas and the MLOps stack canvas, each illustrated by an example. We also discussed their purpose and importance in ML projects. Now, it's your turn, I'd love to hear your thoughts on the canvases; and let me know if you have used it before or if you plan to adopt it in the future!

Oh! I almost forgot, you can download and use the canvases I created from here: the AI canvas, the ML canvas and the MLOps stack canvas.

My aim through my articles is to provide my readers clear, well-organized and easy-to-follow tutorials, offering a solid introduction to the diverse topics I cover and promoting good coding and reasoning skills. I am on a never-ending journey of self-improvement, I share my findings with you through these articles. I, myself, frequently refer to my own articles as valuable resources when needed.

Thanks for reading this article. You can find all the examples of the different tutorials I provide in my GitHub profile. If you appreciate my tutorials, please support me by following me and subscribing to my mailing list. This way, you'll receive notifications about my new articles. If you have any questions or suggestions, feel free to leave a comment.