Infrastructure
MLOps
DevOps practices adapted to the lifecycle of machine learning and AI systems.
MLOps covers everything between a working notebook and a production model: reproducible training, model versioning, deployment, monitoring, drift detection, and retraining pipelines. It is what makes ML systems maintainable at scale.
Modern LLM-based systems have added a new layer often called LLMOps: prompt versioning, evaluation harnesses, cost and latency tracking per prompt, and observability for tool-calling agents.
The tooling landscape spans MLflow, Weights & Biases, LangSmith, Langfuse, Arize, and cloud-native offerings like SageMaker and Vertex AI.
Key points
- Evaluation datasets are the highest-leverage MLOps investment
- Cost per request should be tracked as a first-class metric
- Model or prompt changes need rollback-safe deploys
- Feedback loops (user thumbs-up/down, escalations) drive continuous improvement
Common use cases
Regulated industries requiring auditability
High-volume LLM applications where cost matters
Any system where model behavior changes over time
Frequently asked
Related terms
Applying MLOps in your business?
We help enterprises design, ship, and govern AI systems end-to-end.