Personal data is removed before sending.

The National Design Studio released Rampart, an open-source local masking system using a 14.7 MB ONNX model to block personal data before it reaches AI.

With Rampart, the National Design Studio, an organization affiliated with the White House and tasked with modernizing US federal digital services, publishes an open-source local masking system designed for chatbots and AI forms.

The principle: to detect personal information directly in the browser, then replace it with stable markers before a message is sent to a server or a model provider. The component is based on a small 14.7 MB ONNX model, derived from MiniLM, combined with a layer of deterministic rules for structured elements such as emails, credit cards, IP addresses, or US Social Security numbers. The National Design Studio claims a 98.42% recall rate on an OpenPII dataset covering seven Latin languages, with an announced median latency of 3.9 ms in WebGPU. The technical datasheet, however, remains cautious: non-Latin scripts, certain government identifiers, and circumvention attempts remain documented weaknesses.

Rampart is therefore less of a guarantee of total privacy than a first client-side barrier, useful for preventing sensitive data from leaving the device too quickly.