From a day to ten minutes: reviewing customs entries with AI without the data leaving home

Mexican customs agencies review every customs entry and every invoice by hand. The documents have to be read, quantities, values and codes cross-checked, and any difference caught before it reaches customs. At Griver, that process took a full day.

The documents are also confidential: they hold data about clients, suppliers and goods. Sending them to an external AI service was not an option.

We built an AI OCR API for the RECO prevalidator, Griver's foreign trade platform. The API reads the document, extracts the fields and hands them over ready to cross-check.

The reading is done by a self hosted language model, served with vLLM on GPU in AWS and packaged with Docker. The data never leaves the client's infrastructure and is not stored: each document is processed and discarded.

What took a full day of review now takes ten minutes with the RECO prevalidator. The team spends that time resolving the differences the system flags, not finding them.

The lesson for us was clear: when the data is sensitive, the question is not whether to use AI but where it runs. A model of your own, inside the client's infrastructure, solves both.