Salesforce / EdgeFlow release engineering
AI-assisted guarded releases
for Salesforce Edge.
Let any team member own a release. Make the process safer with AI-assisted checks.
Release ownership was the bottleneck
The Salesforce Edge release process was complex enough that only members of the operations team could run it. A dedicated release owner spent a two-week rotation focused almost entirely on releases.
My primary goal was to make release ownership a team capability: any team member should be able to lead the process with AI assistance. The second goal was to make releases safer. Edge is Salesforce's tier-0 traffic entry point, where a faulty rollout could disrupt Salesforce globally.
A shared, AI-assisted release process
I originated and implemented the release automation as a major use case of EdgeFlow, the general-purpose automation platform I built for engineering operations. The system automated preparation and builds, coordinated staged deployments and kept release status, reports and change-owner testing records together in Slack threads and tracking canvases.
The orchestration covered 46 deployment groups across development, pre-production, canary and production. Separate workflows handled Linux hosts running Docker Compose and public-cloud environments running Kubernetes. A release owner remained accountable for the process, checking that workflows ran and the release progressed correctly.
Safety across an overwhelming amount of information
Each deployment group involved approximately 100 metrics, alongside logs and historical comparisons. Reviewing that volume manually made it easy to miss a metric or a relevant log signal, while comparisons also required keeping previous periods available.
Automation collected the metrics, generated reports and compared post-deployment behavior with immediate pre-deployment, prior-day and prior-week baselines. AI agents assessed release health and whether progression to the next region was appropriate.
AI assessment, staged progression and human change-owner validation remained distinct controls. If PagerDuty incidents appeared after deployment, automation stopped progression and notified the on-call engineer. Recurring checks continued through soak periods, so safety checks were part of the release process rather than a sequence of dashboards someone had to remember to inspect.
The result: shared ownership, safer releases, less manual work
Any team member could now own the release process with AI assistance, rather than relying on an operations specialist. The team gained a safer, repeatable process while retaining a named owner and human checks where required.
Approximately 60 hours → 2 hours of hands-on work per two-week release-owner rotation.
Approximate effort for the person overseeing releases during the rotation.An engineer no longer had to spend the rotation working almost exclusively on releases. The owner oversaw the automated process instead of manually collecting and reviewing its information.
Separately, elapsed rollout time decreased from four weeks to three.