Price the job
Estimate tokens and cost for the number of rows you declare.
ML · ships today
Batch inference
at a known cost
Price the job before the first request, refuse it over budget, record what it actually cost.
Ships as serving/batch-inference
Run a prompt over every row of a dataset against your served model. The cost is checked before the first request, and the real spend is recorded after.
The pipeline
Check the budget
Refuse the job if the estimate exceeds the limit. Nothing is sent.
Run inference
Send each row to your served model or gateway, such as vLLM or LiteLLM.
Record the spend
Price the tokens actually used, next to the estimate.
What the record keeps
The dataset, the prompt, the model it ran against, the approved estimate, the actual spend and the results as a versioned dataset.
No surprise bills
The estimate is checked before the first request is sent.
The cap is the approval
The approved row count is also the hard cap; a larger dataset stops there and the record says so.
Estimate next to actual
Both numbers are kept for every run.
Ships as serving/batch-inference. Model servers and gateways depend on the installed operators.
See how failure can trigger a reviewed improvement cycle →Run it on your own infrastructure.
Apache 2.0 licensed, self-hosted. cairn pipeline run serving/batch-inference