The career path of the MLOps engineer explained
The origin and essence of MLOps
The role of Machine Learning Operations (MLOps) engineer arose from a concrete need within organizations experimenting with artificial intelligence. In the early days of machine learning, data scientists mostly worked in isolated environments, such as local Jupyter Notebooks. They developed mathematical models that performed excellently on historical datasets, but there was no structured method for bringing these models to production environments reliably, scalably, and securely.
Once a model was ready for use, the code was often handed over manually to software engineers. These engineers then had to rewrite the code to make it compatible with the existing IT infrastructure. This process was error-prone, time-consuming, and led to a lack of reproducibility. MLOps engineers fill this gap. They form the operational and infrastructural bridge between data science and traditional software and platform engineering.
The essence of the role lies in automating and standardizing the entire lifecycle of machine learning models. Where DevOps focuses on automating software deployments, MLOps focuses on the trinity of code, data, and model parameters. The goal is to facilitate a continuous stream of model improvements, with every model performing traceably, testably, and consistently in a production environment.
The position in the ecosystem: MLOps versus neighboring roles
Within larger organizations, various specialists work together on data and AI products. Because the boundaries between these roles sometimes blur in job postings, it's important to clearly define the specific focus and the boundary areas. More details on the division of these tasks can be found in the overview of AI roles explained.
The MLOps engineer does not stand alone, but works closely with the following disciplines:
| Role | Primary focus | Typical boundary with MLOps |
|---|---|---|
| Data engineer | Design and maintenance of data lakes, data warehouses, and ETL pipelines. | The data engineer delivers clean, structured data; the MLOps engineer automates its inflow into the training pipeline. |
| ML engineer | Development of model architectures, feature engineering, and model training. | The ML engineer optimizes the model's accuracy; the MLOps engineer ensures the model can be packaged, scaled, and monitored. |
| Platform engineer / DevOps | Generic cloud infrastructure, Kubernetes clusters, network security, and generic CI/CD. | The platform engineer delivers the base infrastructure; the MLOps engineer builds the specific ML tooling on top of it (such as feature stores and model registries). |
| AI engineer | Integration of existing (often external) AI models and APIs into software products. | The AI engineer builds the application layer around a model; the MLOps engineer manages the hosting infrastructure and API gateways of those models. |
| Site Reliability Engineer (SRE) | System availability, uptime, latency, and overall reliability of applications. | The SRE monitors the general infrastructure; the MLOps engineer monitors model-specific performance such as model drift and prediction quality. |
The core tasks per phase in the traditional MLOps lifecycle
The field of work of an MLOps engineer is broad and spans various phases of a model's lifecycle. Below, the main technical responsibilities per phase are described.
1. Training pipelines and orchestration
Manually training a model is unacceptable in a production environment. The MLOps engineer builds and maintains automated training pipelines with orchestrators such as Apache Airflow, Prefect, or Kubeflow Pipelines. These pipelines must be parameterizable and reproducible. This means that when there is a change in the input data or the model configuration, the pipeline can automatically be re-run, going through exactly the same steps.
2. Model registration and version control
Once a model has been successfully trained, it must be stored in a central database, the so-called model registry (for example, MLflow or Weights & Biases). The MLOps engineer ensures that every model is provided with the right metadata. Think of the dataset version used, the hyperparameters, the evaluation metrics (such as accuracy or loss value), and the exact version of the training code. This makes it possible, in the event of problems, to always trace how a model came about.
3. Deployment strategies
Putting a new model into production carries risks. MLOps engineers design and implement deployment strategies to minimize these risks. Commonly used methods are:
- Canary deployments: The new model initially receives a small percentage of live traffic (for example, 5%). Only after a successful evaluation period is this percentage gradually increased to 100%.
- Shadow deployments: The new model receives all live requests in parallel with the old model, but the results are only logged and not returned to the end user. This makes it possible to test the performance and latency in a real production environment without risk to the user.
- Blue-green deployments: Two identical environments (blue and green) are run side by side, allowing routing to be quickly switched from the old to the new version.
4. Monitoring and drift detection
A model that performs well today may be outdated tomorrow due to changing real-world behavior. This phenomenon is known as model drift. We distinguish two forms here: data drift (the input data changes relative to the training data) and concept drift (the statistical relationship between the input and the prediction changes). The MLOps engineer sets up monitoring systems (with tools such as Evidently or Great Expectations) to detect these deviations early and send automatic alerts.
5. Cost control and resource optimization
Training and hosting machine learning models requires significant computing power, often in the form of expensive GPUs. An important task of the MLOps engineer is optimizing these resources. This is done by setting up automatic scaling (autoscaling) based on actual load and selecting the right instance types in the cloud to prevent unnecessary idle time of expensive hardware.
6. Rollback strategies
When monitoring shows that a new model unexpectedly performs poorly or generates errors, the system must be able to intervene immediately. The MLOps engineer configures automatic rollbacks, redirecting traffic straight back to the last known stable model version. This requires the infrastructure to be set up for stateless model microservices that can be started up and torn down quickly.
How LLMs have shifted the field toward LLMOps
The rise of Large Language Models (LLMs) and generative AI has led to a significant shift within the MLOps domain, often referred to as LLMOps. The focus has shifted from training models yourself to effectively managing and optimizing already-existing, gigantic models.
Where previously a lot of time was spent designing custom training architectures, the priority now lies in setting up efficient inference infrastructure. This includes setting up systems for model optimization (such as quantization and distillation) and configuring specialized inference servers such as vLLM or Triton Inference Server. These servers ensure that the latency and the cost per generated token stay as low as possible.
In addition, evaluating generative systems is fundamentally different from evaluating traditional classification models. There is no simple F1 score or MSE metric involved. Evaluation pipelines now have to deal with non-deterministic output. This requires setting up systems that use LLM-as-a-judge methodologies and structurally running regression tests. To set up such evaluation systems, it's necessary to use targeted methods for regression testing for prompts to ensure the consistency of the output.
Prompt management has also become an integral part of the deployment pipeline. Prompts are nowadays treated as software artifacts: they must be tested, versioned, and managed in central repositories. Finally, the complexity of applications that use Retrieval-Augmented Generation (RAG) or multi-agent systems requires a deeper level of tracing. MLOps engineers therefore focus intensively on implementing systems for extensive observability and logging in order to quickly diagnose bottlenecks in complex LLM chains.
Typical entry routes and development paths
There is no uniform training path to becoming an MLOps engineer. Professionals enter the field from various disciplines within IT and data infrastructure. Each route brings specific strengths, but also requires bridging specific knowledge gaps.
From software engineering
Software engineers have a strong foundation in programming principles, CI/CD pipelines, testing, and system architecture. They understand how to write robust code that meets enterprise standards. The main development area for this group is understanding the non-deterministic nature of machine learning. Where traditional software always gives the same output for the same input, that's not the case with ML. They need to delve into the basic principles of data science, model training, and statistics. For this transition, it can be advisable to start with a structured learning path for beginners in machine learning.
From data engineering
Data engineers are used to working with large amounts of data and understand data storage, database optimization, and ETL processes (Extract, Transform, Load) like no one else. They are often already familiar with tools such as Spark or Airflow. To make the switch to MLOps, they need to focus on real-time application architectures (such as exposing models via low-latency APIs), container orchestration with Kubernetes, and the specific lifecycle management tools of machine learning.
From DevOps and systems administration
DevOps engineers have deep knowledge of cloud infrastructure, networking, Kubernetes, and infrastructure-as-code (Terraform). They are excellently equipped to manage hosting environments. Their challenge lies in understanding the specific needs of data scientists. They need to learn to work with data-specific tooling, understand how training pipelines differ from standard software builds, and how to monitor and interpret the performance of mathematical models.
From data science
Data scientists understand the math behind the models, know how to select features, and how to train and evaluate a model. However, they often run into a lack of knowledge in software engineering. To become an MLOps engineer, they need to learn to write clean code, apply version control (Git) professionally, and become skilled in containerization (Docker), CI/CD, and cloud infrastructure.
Essential skills and how to build them independently
Job postings for MLOps engineers often ask for a wide range of tools and technologies. It can be discouraging to build these skills when your current employer doesn't yet apply these methodologies. Still, it's entirely possible to acquire this knowledge independently using open-source software and local development environments.
The most important technological pillars and how to master them independently are:
- Containerization: Learn Docker by packaging your own applications into containers. Make sure you understand how multi-stage builds work to keep container images small and secure.
- Orchestration: Install a local Kubernetes cluster using Minikube or Kind (Kubernetes in Docker). Practice writing Kubernetes manifests for deploying simple web applications and model APIs.
- Experiment Tracking & Model Registry: Run MLflow locally on your computer. Write a simple Python script that trains a model and log the parameters, metrics, and the final model artifact to your local MLflow server.
- Infrastructure as code (IaC): Create a free account with a cloud provider (AWS, GCP, or Azure) and use Terraform to set up and tear down resources. This teaches you how to define infrastructure reproducibly.
- CI/CD: Use GitHub Actions to automatically run tests on your code as soon as you push a commit. Also automate building a Docker image and pushing it to a container registry.
Designing a compelling portfolio project
When applying for an MLOps role, a portfolio with code repositories is often decisive. Employers are not looking for a complicated deep learning model with high accuracy; they are looking for a demonstration of a robust, automated system. A strong portfolio project focuses on reproducibility, automation, and measurability.
An effective setup for a portfolio project consists of the following elements:
- A simple model base: Use a standard dataset (such as house prices or customer churn) and train a simple model (for example, a Random Forest). This keeps the focus on the infrastructure rather than on model optimization.
- Automated training pipeline: Write a script that fetches data, transforms it, trains the model, and saves the model along with its metrics in a registry.
- CI/CD integration: Set up a pipeline that automatically starts when the code changes. This pipeline should run linter checks, run unit tests, and package the trained model in a Docker container.
- API exposure: Develop a simple REST API (for example, with FastAPI) that uses the container to make predictions based on new input data.
- Monitoring and feedback loop: Implement a basic system that logs incoming requests and predictions. Write a script that periodically checks for data drift and sends a notification if this is the case.
Document the project thoroughly in a README file. Explain how the architecture is put together, which design decisions you made, and how someone can reproduce the project locally with a minimal number of commands. More tips on structuring such projects can be found in the guide on portfolio projects in practice.
Growth directions and specializations
The career path of an MLOps engineer offers various opportunities for specialization and growth as one gains more practical experience.
Senior and Lead MLOps Engineer
In this role, the focus shifts from building individual pipelines to designing the overall technical strategy for the organization's AI platform. You select the tooling stack, define security standards, and ensure integration with existing enterprise systems. In addition, you guide and coach junior team members.
Platform Architect (AI & Data)
As a platform architect, you are responsible for the bigger picture. You design the data infrastructure that supports data engineering, data science, and MLOps alike. This requires a deep understanding of cloud architecture, data modeling, network security, and compliance legislation.
FinOps Specialist for Machine Learning
Given the high costs associated with training and running large models (LLMs in particular), some MLOps engineers focus entirely on cost optimization. They analyze resource consumption, implement advanced autoscaling strategies, and advise on the most cost-efficient ways to meet performance requirements.
The Dutch MLOps market
In the Netherlands, demand for MLOps engineers is concentrated in specific sectors that have mature data teams and the resources to implement AI solutions at scale. This influences the type of work and the technological challenges you encounter.
The main employers can be found in the following segments:
- Scale-ups and tech companies: These organizations are often at the forefront of adopting modern open-source tools and cloud-native architectures. The focus here is on speed, flexibility, and quickly bringing new features into production.
- Financial sector (banks and insurers): Within this sector, security, traceability, and compliance are of the utmost importance. MLOps engineers work here in strictly regulated environments and spend a lot of time setting up audit trails, data lineage, and bias detection to comply with laws and regulations such as the European AI Act.
- Government and semi-public sector: Here, the emphasis is strongly on transparency, model explainability, and data sovereignty. This means work is often done with on-premise infrastructures or specific European cloud solutions, which brings unique challenges for setting up standard systems.
- Specialized consultancies: Many organizations don't have the in-house knowledge to set up an MLOps platform themselves and hire external expertise for this. As a consultant, you work on diverse projects for different clients, which allows you to quickly build up varied practical experience.


