Overview ZenML is positioned as a unified infrastructure layer for production AI, provided by ZenML Labs. Its core consists of two parts: ZenML is responsible for orchestrating pipelines and agents on the infrastructure of your choice; Kitaru is responsible for replaying them on production data. The official emphasis is that both projects are always open source.
Purpose ZenML is aimed at teams that want to "deliver trustworthy AI," emphasizing building first, then proving. It is used to orchestrate AI pipelines and agents in production environments and, before changes go live, evaluate the effects of changes through replay on production data.
Main capabilities
- Orchestration on infrastructure of your choice: Write pipelines and agents in Python and run them on your existing orchestrators and cloud platforms, migrating from notebooks to Kubernetes without rewriting.
- Production data replay: Import sessions in which agents actually ran, identify recurring patterns, and replay a change against those sessions. Before going live, see which metrics improve, which regress, and how costs change.
- Open source and self-hosting: Both ZenML and Kitaru are licensed under Apache 2.0, can be self-hosted, allow reading the source code, and keep data in place.
- Enterprise-grade compliance: SOC 2 Type II and ISO 27001, running on your own infrastructure.
How to use According to the official materials, the methods include: writing pipelines and agents in Python; running them on existing orchestrators and clouds; importing real agent run sessions as test sets; and replaying changes to compare improvements, regressions, and costs. More specific installation commands, code examples, or operating steps are not yet provided in the materials.
Use cases Suitable for teams and organizations that need to deploy machine learning or AI pipelines and agents to production and want to validate the effects of changes with real data before going live. Official customer cases involve scenarios such as JetBrains centralizing ML on Kubernetes, ADEO Leroy Merlin reducing time to market from 2 months to 2 weeks, Brevo accelerating model development by 80%, and Cross Screen Media training models for 210 markets within hours.
Limitations and information not yet provided The materials do not provide specific pricing and fee information, installation commands, code examples, a complete list of supported platforms or orchestrators, performance metrics, or detailed usage steps. The above content is not yet provided.
