Atlassian Builds ML Studio To Power Rovo Workflows

Atlassian published a detailed engineering writeup on ML Studio, its internal ML development platform, which the company says powers Rovo Search and Chat, the Teamwork Graph, and Confluence AI features, running thousands of production workflows daily. Atlassian reports the platform serves more than 5 million monthly active Rovo users, has generated over 900,000 access-controlled datasets, and supports roughly 120,000 monthly workflow runs across 100-plus ML teams, alongside about 20,000 monthly model iterations and more than 2,000 reusable ML modules with over 200,000 monthly iterations.

The platform rests on three pillars, per the post: composable, versioned ML modules managed through a Git-backed module layer; a Workflow Orchestrator with a CLI and portal that schedules jobs across platforms like Databricks and supports hot clusters, nested workflows, and CRON scheduling; and a multi-layer compliance framework that enforces access control at the user, domain, and column level, with data-classification tags that propagate automatically from input to output tables. Atlassian's authors, Apoorva Uday Nayak, Longfei Zhang and Gaurav Awadhwal of the AI & ML Platform team, describe the architecture as a reusable pattern beyond Atlassian's own use.

The most concrete numbers in the post are productivity metrics: Atlassian says local development builds cut Python module build times from minutes to under 30 seconds, with similar gains for Docker builds; about 80% of ML Studio workflows use automatic result caching, saving a reported 1,000-plus hours of workflow execution time per month; and PR-free experimental workflows now account for more than half of all runs, saving an average of 100-plus hours per day across Atlassian's Central AI org. These are self-reported, internal figures rather than independently benchmarked results, but they give teams designing similar in-house ML platforms concrete targets, caching hit rates, build-time reductions, and experimentation-to-production friction, to compare against.

Watch whether Atlassian open-sources or documents any of ML Studio's components (module management, the workflow orchestrator, or the column-level classification framework) the way some hyperscalers have with internal MLOps tooling, and whether Atlassian publishes follow-up posts with cost or GPU-utilization figures, which this post does not include despite describing GPU cluster integration for distributed training.