Service
The application around a model
A model in a notebook answers a question once. A model in production answers it all day, for people who never saw the notebook. We build the part in between: the service, the tests, the deployment and the monitoring.
In three steps
How a model reaches production
Serve
The model behind a clear interface: an API or a scheduled job, with input checks and a known answer when something is wrong.
Release
Tests run on every change. A release is one repeatable command, and so is going back to the last good version.
Monitor
Inputs, outputs, speed and errors are watched. An alert reaches a named person before users notice.
01 Who it is for
Teams with a model that works in a notebook or a prototype and now has to serve real users, other systems or a daily report. It fits best when the person who built the model is also the only person who can run it.
02 Signs you need it
- Running the model means one person typing commands by hand.
- A release is a careful afternoon, and going back is harder.
- Nobody can say whether live predictions still match what was tested.
- Nobody can say which version of the model is live, or which data trained it.
- Users find failures before the team does.
03 What you get
- A service or a job
- The model behind a defined interface, with input checks and clear errors.
- Tests
- For code, for data and for model quality, run on every change.
- Repeatable releases
- One command to release, one to roll back, and each model version traced to its code and data.
- Monitoring and alerts
- Inputs, outputs, speed and errors, with alerts sent to a named person.
- A dashboard
- Where the people who rely on the model can see how it is doing.
- A runbook
- How to release, roll back and recover, written for the next engineer.
All of it is code in your repository, running in your accounts.
04 What we need from you
- The model, or the notebook and the data it was trained on.
- The place it must run: a cloud account, a server or a cluster.
- Who will use it, how often, and how fast an answer must be.
- One person who can say when a prediction is wrong.
05 What a first pilot looks like
One model behind one interface in your account, with tests, a release you can repeat, and the first monitoring and alerts. You can run each week's version yourself.
The scope, the price and the length are agreed in writing before it starts. How a pilot works.
06 When we would say no
- The model is not good enough yet. Putting it in production would only make its mistakes faster, and we will say so.
- A managed platform you already pay for covers the need.
- The main need is a large product interface. The application around the model is our part; a full product team is not.
Next
A model that only one person can run?
Tell us what the model does, where it runs today, and who needs it. An engineer replies.