Union.ai grew out of the Flyte project, the open-source workflow engine used widely for data and ML pipelines, and it keeps that lineage in its design: you write ordinary Python tasks and task environments, and the runtime takes care of scheduling, caching, retries and typed data passing between steps. What Union adds on top is durability. If a run dies because a node disappears, a GPU is reclaimed, an API returns a 429 or an out-of-memory kill fires, the run is recovered and continues from the failed step rather than the beginning of the pipeline.
The current platform is organised around a handful of capabilities. The AI factory manages named, versioned artifacts so a checkpoint or dataset has a traceable producer and consumers, which matters for reproducibility. Production agents let teams keep the agent framework they already use, including OpenAI, Claude, PydanticAI and CrewAI, while making every turn of the loop durable, traced and sandboxed. The inference gateway puts external model APIs and self-hosted vLLM, SGLang or Ollama endpoints behind a single governed interface with virtual keys, budgets and guardrails. Around these sit scheduling primitives such as priority queues, concurrency limits, quotas and cluster selectors.
Deployment is deliberately bring-your-own-cloud. Union hosts the control plane, which stores references and metadata rather than payloads, while your data, models and compute stay inside your own perimeter, so your negotiated cloud rates and credits still apply. Union is available on AWS Marketplace and through self-service signup, with an optional managed-cluster service billed per cluster.
