A four-person research team at a European genomics lab had a model that worked beautifully — in a notebook, on one GPU, run by one person who knew exactly which cells to execute in which order. Turning that into something their collaborators across three countries could rely on felt like a quarter of engineering work. It took a weekend.
This is the story of how they got from prototype to production without hiring an infrastructure team, and what they learned along the way.
Friday: containerise and deploy
The first step was the least glamorous: pulling the notebook apart into a small inference module with a clean function signature. From there, our CLI packaged it into a container and pushed it to a GPU-backed endpoint. By Friday evening they had a URL that returned predictions, autoscaled from zero, and cost nothing while idle.
No Kubernetes manifests, no load balancer configuration, no capacity planning. The platform handled placement and scaling; the team handled the science.
Saturday: multi-region and real traffic
With collaborators in Madrid, Frankfurt and Stockholm, latency mattered. Flipping the endpoint to multi-region was a one-line change, and requests began routing to the nearest healthy replica automatically. By Sunday the service was handling real traffic from three sites, with tracing and per-request cost visible in the dashboard.
The lesson they took away wasn’t about any single feature. It was that the gap between “it works on my machine” and “it works for everyone, all the time” had quietly shrunk from months to a weekend.