ZenML GmbH
ZenML builds open-source infrastructure for orchestrating production AI workflows: Python pipelines and agents can run across the infrastructure a team already uses, from a laptop to Kubernetes and major clouds. Its newer Kitaru project replays real production agent sessions as repeatable evaluations before changes ship.
At a Glance
- ML and AI engineering teams
- Data-science teams moving models to production
- Enterprises with multi-cloud or Kubernetes infrastructure
- Organizations building LLM and agent applications
- +1 more
AI Tools by ZenML GmbH
(1)
ZenML
Open Source ML Pipeline Platform
Discussions
No discussions yet
Be the first to start a discussion about ZenML GmbH
Latest News
ZenML/Kitaru published its evaluation of sandbox runtimes for long-lived coding agents and selected Daytona as its current choice.
Introduced the new Kitaru, turning production traces into replayable test scenarios, expert-reviewed cohorts and evaluators.
Published a robotics-focused MLOps announcement about keeping robot-learning loops reproducible across distributed GPUs and clusters.
Announced Kitaru's durable runtime around Claude Agent SDK invocations, including checkpoints, artifacts, replay boundaries and waits.
Products & Services
Apache 2.0 Python framework for portable, reproducible ML and AI pipelines; tracks runs, artifacts, metadata and lineage while abstracting infrastructure.
Managed control plane for production teams, adding model and artifact control planes, snapshots, Codespaces, governance, SSO/RBAC and enterprise deployment options.
MCP integration that lets compatible assistants such as Claude Desktop and Cursor query pipeline runs, analyze metadata and trigger deployments using natural language.
Skills for coding agents, including a Quick Wins skill that audits a ZenML codebase and implements best-practice MLOps improvements.
Market Position
ZenML positions itself as an open-source, cloud- and tooling-agnostic AI orchestration layer that sits between Python workflow code and infrastructure, emphasizing portability, composable stacks and avoiding vendor lock-in. It differentiates from Airflow, Luigi and Prefect by modeling ML artifacts, lineage, caching and model workflows directly; Kitaru competes in replay/evaluation and durable-execution tooling alongside Temporal, Restate, DBOS, Inngest, Convex and related agent-runtime products.
Leadership
Founders
Adam Probst
Co-founder and CEO. Before ZenML, he co-founded MAIoT (2017–2021) and worked at Knorr-Bremse in industrial engineering and operational excellence; the founders' prior startup built machine-learning solutions for predictive maintenance in commercial vehicles.
Hamza Tahir
Co-founder and CTO. Previously co-founder and CTO of MAIoT; he and Probst had first-hand experience productionizing machine learning for predictive-maintenance projects and the resulting hand-off and reproducibility problems.
Executive Team
Adam Probst
Co-founder & CEO
Former co-founder of MAIoT and industrial engineering/operational-excellence professional at Knorr-Bremse.
Hamza Tahir
Co-founder & CTO
Former co-founder and CTO of MAIoT; leads the technical direction of ZenML and Kitaru.
Founding Story
Probst and Tahir started ZenML after seeing the same problem repeatedly at their first startup: ML pipelines for commercial-vehicle predictive maintenance had to be adapted for different legacy tools, IT processes and modalities, while hand-offs between data science, ML engineering and DevOps created an ownership dilemma. They envisioned a Python, workflow-centered, infrastructure-agnostic layer that lets data scientists own a pipeline through production without needing to master every deployment system.
Business Model
Revenue Model
Open-source self-hosting is free; recurring SaaS subscriptions monetize the hosted ZenML Pro/Cloud control plane and Kitaru Cloud, while enterprise customers purchase custom plans with governance, support and deployment capabilities.
Pricing Tiers
Self-hosted; unlimited executions and projects; pipeline/flow orchestration, artifact management, basic model registry and community support.
SaaS plan for production teams; example tier is 2,000 executions, 3 projects and 5 snapshots, plus Model/Artifact Control Planes and Codespaces.
Unlimited executions/projects, SSO, custom RBAC, audit logs and air-gapped deployment.
Self-hosted Apache 2.0 version with unlimited import/recording and replay on the customer's workers.
Hosted plan with 3 agents, 2 seats, 90-day session retention and included replays/experiments; 14-day full-access trial.
Unlimited agents, custom seats/retention, SSO, audit logs and remote worker pools.
Target Markets
- ML and AI engineering teams
- Data-science teams moving models to production
- Enterprises with multi-cloud or Kubernetes infrastructure
- Organizations building LLM and agent applications
- Research institutions, universities and early-stage ML product companies
- Productionizing classical machine-learning training and batch-inference pipelines
- LLMOps, RAG and generative-AI workflows
- Building and deploying AI agents on existing infrastructure
- Reproducible experimentation, model/version governance and CI/CD
- Robotics training and evaluation loops spanning cloud, cluster, simulation and edge hardware
- Replay-based regression testing and cost/quality evaluation of production agents
- JetBrains
- ADEO / Leroy Merlin
- Brevo
- AXA