- Home
- AI development
- MLOps
MLOps
MLOps: keep an AI system working after launch, and know what it costs
MLOps is the routine that keeps an AI system working once people depend on it. Model versions retire, prices change, answers drift, and an API key with no spending cap can run up a bill overnight. We host and monitor the system, log every model call, set spending caps and keep a way to roll back.
- Prices published on the site
- First working version in two weeks
- Support month to month, cancel any time
An AI system that passed its tests in week one keeps changing underneath
The model provider updates or retires a model version, and the answers shift. Prices change. Your documents and your customers' wording move away from what the system was tested on. A script that loops, or an API key with no spending cap, can burn through a month's budget in a day.
Operations covers four things: a controlled release, a record of every model call, a weekly check against the test set, and a safe way to update. It is routine work, and we do it as part of support for systems we build or take over.
Release, watch, review, update, and around again
Release
Version the prompt, the model and the settings together. Switch the new version on gradually, for a share of requests or one team first.
Watch
Cost per task, errors, response time and a daily sample of outputs that a person can read.
Review
A weekly sample scored against the test set. Alerts fire on a spike in cost or errors.
Update
A new model version or a fix goes through the same tests. If the scores are worse, we roll back to the last version that passed.
On day one we put in the call log and the spending cap. The rest grows with the system and the number of people using it.
What is in place once a system is in production
Hosting
The system runs on our own servers and Cloudflare, or in your cloud account. Accounts and code repositories are in your name.
A log of every model call
Which model, which prompt version, how many tokens, how long, what it cost and what came back. Personal data is masked in the log where it must not be stored.
Spending caps
A monthly spending limit for each system, set in the provider account where it offers one and in the application where it does not, with an alert before the limit is reached.
A rollback
The last version that passed the tests stays ready to deploy. Going back is a switch for our code and settings. A model the provider has retired cannot come back.
Scheduled checks
A daily sample and a weekly score land in the inbox of the person who owns the system, in a short report.
Provider changes
We follow the model retirements and price changes announced for the models your system calls, and test the replacement before the old one goes away.
How much operations a system needs
Set up the full routine when
- The system answers customers, or touches records, orders or email
- More than a handful of people use it every day
- It calls a paid model API, so a bug or a loop costs real money
- It will run for more than a few months and go through at least one model change
Keep it small when
- A pilot used by two people needs a call log and a spending cap, and a full pipeline would cost more than it protects
- A scheduled script whose output a person reads every day anyway
- A tool whose answers are reviewed every time: watch the bill and skip the dashboards
Setup once, then a support plan
One workflow running end to end, or one internal tool.
Several systems made to work as one: integrations, roles, reports.
Setup for one system, with the call log, the caps, the weekly test and the rollback, usually lands in one process. Several systems under one routine land in connected systems.
The ongoing price is the support plan. Care covers monitoring, small fixes, updates and incident response. Active adds continuing improvements and new scenarios.
Care $300 to $500 a month. Active $750 to $1,500 a month. AI model usage, GPU time and hardware are paid to the provider, on accounts in your name.
Questions about running AI in production
What is MLOps, in plain words?
MLOps is the work of keeping an AI system running after launch: releasing changes safely, watching cost and errors, testing new model versions and rolling back when something gets worse. For a business system that calls a model through an API, it comes down to logging, spending caps, a test set and a rollback.
Why would a model that works today stop working?
The provider changes it. Versions are updated and retired on the provider's schedule, answers shift a little with each one, and prices change. Your own data drifts too: new product names, a new form layout, customers who write differently. A weekly score against the test set usually shows the change early.
How do you stop a runaway API bill?
Each system gets its own API key with a monthly cap, and the log shows the cost of every call, so a spike can be traced to the task or loop that caused it. Keys are kept outside the code. A looping script hits the limit, and a leaked key is revoked and replaced; the log shows which system it belonged to.
Where does the system run?
On our own servers and Cloudflare, or in your cloud account. Either way the accounts and code repositories are in your company's name, and the system can run on your own server. If it calls an outside AI model, we tell you which one and what data it sees.
Do we need this for a small pilot?
A pilot needs a log of model calls and a spending cap, and we set both up in the first week. The weekly test and the gradual release start to matter when more people depend on the system.
What does the support plan cover?
Care covers monitoring, small fixes, updates and incident response. Active adds continuing improvements and new scenarios. Both run month to month and you can cancel any time, with the paid month running to its end. Pricing lists the rates.
Related services
- LLM evaluation Test which model does your job best
- AI agents Agents that finish a task, with your sign-off
- Servers and support Servers, backups, monitoring, monthly support
- Local AI setup AI on hardware you own, set up on-site
- LLM fine-tuning Train a model on your own examples
- AI development Every AI service we build, on one page
Tell us what's slowing you down.
You get a free written review by email within two business days.