MLflow

Experiment tracking was MLflow's original job: log parameters, metrics and artifacts from each training run, compare runs in a web UI, and promote models through a registry before serving them on Docker, Kubernetes, SageMaker or Azure ML. A large GenAI layer came later. Tracing built on OpenTelemetry records each LLM call, tool call and retrieval step in an agent, with one-line autologging for frameworks such as LangChain, DSPy, CrewAI, LlamaIndex and the OpenAI SDK. On top of traces sit evaluation with built-in LLM-judge scorers or custom metrics and a versioned prompt registry that can also tune prompts automatically. A gateway component puts many model providers behind one OpenAI-compatible endpoint with rate limits, fallbacks and traffic splitting. Clients exist for Python, TypeScript and Java. Getting started takes one command, uvx mlflow server, which serves the UI on port 5000; team setups move metadata into PostgreSQL or MySQL and artifacts into S3-style storage, or use a Databricks workspace. It is Apache 2.0 and free to self-host commercially. Suits ML teams wanting one system for model lifecycle and LLM observability.