A Quiet Quest to Tame AI’s Production Disruption
In the high-stakes world of enterprise artificial intelligence, there are specters that haunt boardrooms and engineering teams alike. It’s a prototype that never leaves the lab. It’s common for great machine learning models to run with amazing accuracy on data scientists’ laptops to fail in the vast gulf between development and operations. This is not an isolated issue. It’s an industry-wide trend. According to a widely cited report detailed by VentureBeat, a staggering 87% of data science projects never make it into production. This gap represents billions in unrealized value and countless hours of wasted effort.
The field of MLOps (Machine Learning Operations) has emerged as the prescribed cure for this disease, promising to blend the agile, iterative world of data science with the robust, scalable principles of DevOps. However, for many organizations, this treatment brought its own complex side effects. The very tools designed to streamline processes, such as the powerful but notoriously complex Kubeflow, often require data scientists to become experts in cloud-native infrastructure, a world of containers, clusters, and YAML files far removed from their core expertise in statistics and algorithms. In response to this friction, a new class of tools is emerging that focuses on radically simplifying the path to operations rather than adding functionality.
One of the most exciting efforts in this area, to say the least, is an open source project aptly named ‘kaos’. Billed as a “platform for coordinating and managing the machine learning lifecycle,” the platform was developed by Alejandro Saucedo, a prominent figure in the AI field known for his work as director of engineering at MLOps company Seldon and principal investigator at the Ethics AI and Machine Learning Institute. The project, available on GitHub, is an incisive and unique toolkit born not so much from a vast platform as from the very real problems of deploying ML at scale. Its central theme is that the power of modern infrastructure like Kubernetes should be an invisible engine, not an obstacle, for scientists building models.
Building a command line bridge across difficult waters
Essentially, “kaos” is an exercise in abstraction. We know that the cognitive load required to manage a full-scale Kubernetes deployment is a major reason why ML models stall. Data scientists who specialize in feature engineering and model tuning are suddenly asked to become cloud infrastructure engineers. The “kaos” framework attempts to solve this problem by providing a simple and intuitive command line interface (CLI) that acts as a bridge between a data scientist’s local environment and the complex distributed environment of a production cluster. The workflow boils down to a few simple commands: “kaos build”, “kaos train”, “kaos run”, and “kaosserve”.
This approach allows data scientists to package code and dependencies into a reproducible format (“build”), run training jobs on scalable remote infrastructure (“train”), and deploy the resulting models as production-ready API endpoints (“serve”) without writing a single line of Kubernetes configuration. This philosophy directly reflects the “developer experience” revolution seen in traditional software engineering, where platforms like Heroku have gained many fans by abstracting server management. The “kaos” documentation explicitly cites this similarity and aims to provide a “Heraku-like experience” for machine learning, a powerful and resonant goal for anyone getting lost in the weeds of infrastructure management.
Infrastructure as code: The foundation for reproducibility
This simplicity is supported by strict adherence to Infrastructure as Code (IaC) principles. All components of an ML environment, from data sources and code to specific hardware and library versions, are defined in versioned configuration files. This is very different from ad-hoc, notebook-driven experiments that often result in “it worked on my machine” issues. By enforcing this structure, “kaos” ensures that all training runs and model deployments are fully reproducible, auditable, and transparent. In theory, an engineer could check out a specific commit from a repository and completely recreate the exact model and environment that existed months or even years ago.
This disciplined approach is fundamental to building trust in AI systems, especially in regulated industries such as finance and healthcare, where model lineage and auditability are non-negotiable. By incorporating IaC principles directly into a data scientist’s workflow, “kaos” not only facilitates deployment; It will be more robust and reliable. Treat the entire ML lifecycle as one cohesive software project, following the same standards of version control, testing, and automation that govern traditional application development.
Navigating the crowded field of MLOps tools
There is no shortage of tools in the MLOps domain, and large, comprehensive platforms are vying for supremacy. Chief among them is Kubeflow, a project aimed at providing end-to-end ML toolkits directly on Kubernetes. According to the official documentation, the goal of Kubeflow is to make it “simple, portable, and scalable” to deploy ML workflows on Kubernetes, as explained on the Kubeflow project website. However, its comprehensive nature makes it significantly more complex to operate. Installing and managing Kubeflow instances is a significant undertaking and often requires a dedicated platform engineering team.
“kaos” positions itself not as a direct competitor to these giants, but as a lightweight, opinionated alternative. It is designed for teams who feel Kubeflow’s comprehensive nature is overkill for their needs. Rather than trying to be everything to everyone, “kaos” is focused on doing one thing very well: simplifying the core loop of training and serving models on top of your existing Kubernetes infrastructure. It’s not reinventing the wheel. It builds on the fundamental power of Docker and Kubernetes and provides a user-friendly facade that unlocks its capabilities to a wider audience. This focus makes it an attractive option for smaller teams and organizations looking for a more gradual and less disruptive entry into MLOps.
An idea ahead of its time?
Developed by practitioners with deep industry experience (Saucedo is a visiting scholar at the prestigious Alan Turing Institute in addition to his other roles), “kaos” represents a clear vision for a more accessible future for MLOps. Development activity on this project appears to have peaked in 2021 and 2022, and it currently exists as a mature proof of concept rather than a commercially backed, rapidly evolving product. However, its core idea is more realistic than ever. The industry continues to grapple with the same complexity and usability challenges that “kaos” was designed to solve.
The ultimate success of specialized tools like “kaos” will depend on community adoption and the willingness of organizations to adopt a more modular, best-of-breed approach to their MLOps stacks rather than committing to a single monolithic platform. While the project itself may not be at the center of today’s MLOps conversation, its philosophy is reflected in a new wave of tools focused on the developer experience. A persistent challenge for the AI industry is finding the right balance: providing a powerful, scalable infrastructure without requiring data scientists to become mechanics for the systems they use. The quiet and thoughtful design of ‘kaos’ serves as a valuable blueprint in its ongoing exploration.
