Ramp’s ML platform team rebuilt their orchestration layer on Prefect in a single quarter. They migrated 200 flows off Metaflow, grew to 70+ active contributors, and gave new users a 30-minute path from idea to production.
Ramp on Prefect
Ramp moves fast. We need partners who do the same. We’re making a bet on not just the tool as it is today, but also where it’s going.
Ryne CarboneStaff Machine Learning Engineer, RampWatch the story
Two minutes on why the team chose Prefect and what changed.
Prefect is very open-ended, which is great. Being opinionated matters, and it should be the company or your ML platform team doing that. Not the framework.
Ryne CarboneStaff Machine Learning Engineer, RampFrom Metaflow to Prefect
When Metaflow’s opinionated model started getting in the way, Ramp’s ML platform team rebuilt their orchestration layer on Prefect. Three months later, they had ported every flow, and their users had already added more on top.
Before
200
Metaflow flows
After
350
Prefect flows, one quarter later
How Ramp uses Prefect
One platform for technical and non-technical contributors, and agents that can debug their own flows.
Python decorators
Ramp's developers add @flow or @task to Python they're already writing and the code is ready to deploy. Going from prototype to production didn't require much refactoring, and Prefect stayed out of the way of testing and running workflows.
Templating and event triggers
Ramp's daily ML batch predict is a single Python flow with Prefect decorators. Users configure their own input parameters, event triggers, schedules, and resources, then deploy variants separately while the platform team maintains the core.
CLI, API, and skills
Ramp connects its coding agents to the Prefect CLI, API, and skills to investigate infrastructure and flow failures. Their Automated Debugger watches the alerts channel, investigates failures through an internal agent harness, and drafts a PR for review.
On developer experience
Prefect felt mostly like writing Python code. Going from prototype to production didn’t really involve much refactoring.
Ryne CarboneStaff Machine Learning Engineer, RampInfrastructure decorators
There’s two fronts where infrastructure decorators help, and they’re both during development. They bridge the gap between development and production, and they unlock extra compute, bigger machines than your laptop has.
Six months in
In our ML platform repo, we’ve had more commits in the past six months than our entire history before that.
Ryne CarboneStaff Machine Learning Engineer, RampPrefect gives platform teams the flexibility to set their own patterns and gives every contributor a short path from idea to production.