Dev/test cycle time, with workflows moving from local to CPU to full GPU cluster with no image rebuilds
Time to secure short-term GPU reservations on AWS, down from months of advance planning
WHOOP is a 24/7 wearable and health platform that measures sleep, strain, recovery, and heart health, turning that data into personalized daily guidance members can act on. Behind that experience is a small MLOps team supporting a much larger group of data scientists, all racing to ship models that make WHOOP smarter about the people who wear it. To keep pace, the WHOOP MLOps team rebuilt its machine learning infrastructure on the Anaconda Platform, using AI Orchestration to move from an in-house system that had served them well to a governed, multi-cloud foundation that scientists and engineers now use every day.
The Challenge
The WHOOP MLOps team, led by Staff Software Engineer & Team Lead Adam Berger, had built an in-house tool on Amazon Sagemaker to handle data analysis and model training. It worked well enough to get the company through its first several years of shipping models, but it couldn’t keep up with a science team that wanted to move faster and try new things.
Adding new capabilities meant fighting the architecture rather than building on it. Defining a workload’s logical flow entirely as code, so both the infrastructure and the underlying science could be versioned together, would have required a major re-architecture the team didn’t have time for.
The day-to-day cost showed up in the dev/test loop. Testing a one-line fix could take several minutes. Forget to add a dependency, and the fix was to rebuild the entire container image, another several-minute wait. For a team trying to help data scientists iterate quickly, the platform itself had become the bottleneck.
Approach
WHOOP moved its ML workloads onto the Anaconda Platform, adopting AI Orchestration, powered by Metaflow, as the operational layer connecting experimentation to production. The team extended the Metaflow and AI Orchestration SDKs with WHOOP-specific business logic, making it easy for scientists to read and write highly partitioned datasets, and standardized how the company logs models and training metrics across open source tools like TensorBoard and MLflow.
Governance came built in rather than bolted on. AI Orchestration let the team segment workloads by data sensitivity and by compute type, so users without the right access can’t reach sensitive data, and a small job can’t accidentally land on an expensive GPU cluster meant for something else. As Adam put it, “the goal of any platform team is to make it easy for customers to do the right thing and make it hard for them to do the wrong thing.”
Compute flexibility mattered just as much. WHOOP workloads run on AWS, and GPU capacity is scarce and hard to plan for months out. AI Orchestration gave the team access to GPU clusters across hyperscalers and neoclouds through a single pane of glass, so scientists can point a configuration at a different provider without rewriting scripts. And during peak periods of unplanned GPU usage, the AI Orchestration platform simply queues jobs until one becomes available.
The clearest example of the Anaconda-AWS relationship: the WHOOP MLOps team worked directly with Anaconda to build a feature on top of AWS Capacity Blocks for ML, making short-term reservations available in the platform within minutes. That gives scientists GPU access with a day’s notice instead of months of planning, a problem Anaconda and WHOOP solved together.
Underneath it all, WHOOP uses Conda to manage Python environments project by project, and points IDEs like VS Code and Cursor directly at those environments so they’re ready the moment a developer opens their editor.
Transforming from Individual Capability to Team Capacity
WHOOP moved its ML workloads onto the Anaconda Platform, adopting AI Orchestration, powered by Metaflow, as the operational layer connecting experimentation to production. The team extended the Metaflow and AI Orchestration SDKs with WHOOP-specific business logic, making it easy for scientists to read and write highly partitioned datasets, and standardized how the company logs models and training metrics across open source tools like TensorBoard and MLflow.
Governance came built in rather than bolted on. AI Orchestration let the team segment workloads by data sensitivity and by compute type, so users without the right access can’t reach sensitive data, and a small job can’t accidentally land on an expensive GPU cluster meant for something else. As Adam put it, “the goal of any platform team is to make it easy for customers to do the right thing and make it hard for them to do the wrong thing.”
Compute flexibility mattered just as much. WHOOP workloads run on AWS, and GPU capacity is scarce and hard to plan for months out. AI Orchestration gave the team access to GPU clusters across hyperscalers and neoclouds through a single pane of glass, so scientists can point a configuration at a different provider without rewriting scripts. And during peak periods of unplanned GPU usage, the AI Orchestration platform simply queues jobs until one becomes available.
The clearest example of the Anaconda-AWS relationship: the WHOOP MLOps team worked directly with Anaconda to build a feature on top of AWS Capacity Blocks for ML, making short-term reservations available in the platform within minutes. That gives scientists GPU access with a day’s notice instead of months of planning, a problem Anaconda and WHOOP solved together.
Underneath it all, WHOOP uses Conda to manage Python environments project by project, and points IDEs like VS Code and Cursor directly at those environments so they’re ready the moment a developer opens their editor.
Results
With the Anaconda Platform in place, the WHOOP MLOps team moved from an in-house system that had served them well but couldn’t keep pace with a growing science team, to a faster, governed foundation that’s already reshaping how the company builds and ships models:
- 90+ WHOOP employees now use the Anaconda Platform, including data scientists, software engineers building on AI models, and analysts reviewing outputs.
- Seconds, not Minutes. Testing a change now works like running it locally, and scientists can move a workflow from local to CPU to a full GPU cluster in a matter of seconds, no image rebuilds required.
- GPU access without re-architecture. WHOOP runs its GPU workloads primarily on AWS, with neocloud providers as overflow capacity when demand outstrips availability, and can secure short-term AWS GPU reservations with about a day’s notice instead of months of planning.
- Growing AWS footprint. WHOOP has gone from running occasional, ad hoc jobs on on-demand compute to continuously running jobs on a multi-GPU cluster on AWS, a shift the team credits directly to AI Orchestration.
- Daily manual reporting became a hosted application. The WHOOP signal processing team used to test new hardware materials, generate a report, and email it out, then rerun the same analysis the next day. They rebuilt it as a dynamic, hosted application on AWS within the Anaconda Platform that updates on a regular cadence and includes toggles so stakeholders can view different time periods or past versions on their own.
- Full lineage on every run. The Anaconda Platform tracks the exact code, data, and dependency versions behind every model, making it possible to reproduce a result from last week or last year.
“It’s been really exciting to use Anaconda to spin up data science and machine learning workloads that we could have never dreamt about running a year ago. I’ve also had a really good time working with the technical team at Anaconda to build new features into the platform.” — Adam Berger, Staff Software Engineer & MLOps Team Lead, WHOOP
Better Together with AWS
The WHOOP story is also an AWS story. Anaconda’s orchestration and AWS’s compute depth solved together what neither could have solved alone: day’s-notice GPU access, governed workload segmentation, and neocloud failover, all without re-architecting a single script.
Looking Forward
WHOOP has spent the past year training, multimodal foundational models on the Anaconda Platform. The next step is production: bringing those models to WHOOP members in ways the team expects will improve core features, while unlocking entirely new insights about how members’ bodies and daily behavior connect.
The MLOps team’s advice to others starting a similar migration: talk to the data scientists using your current platform before you build the next one. WHOOP chose Anaconda after sitting down with its own science team to understand what was and wasn’t working, and Adam points to a broader lesson for any platform team operating in this space: “data scientists and software engineers have really different development cycles,” and understanding that difference is what makes a platform genuinely useful rather than just technically capable.
Ready to accelerate your AI initiatives with trusted open source? Discover how the Anaconda Platform can transform your data science workflows while maintaining enterprise-grade security and governance. Contact our team for a personalized demonstration.
Anaconda Platform
The Only Unified Platform for Open Source