Interview

National Design Studio launches Ramparts, a 15MB browser-native PII redaction model that runs entirely on-device

Jun 29, 2026 with Edward Coristine & Tiger

Key Points

  • National Design Studio launches Ramparts, a 15MB browser-native PII redaction model that runs entirely on-device without server calls, eliminating the need for users to trust external privacy claims.
  • The team evaluated 72 base models before settling on a fine-tuned MiniLM architecture, designed to run on legacy Android and iOS hardware where competing models start at 50MB.
  • Ramparts targets government forms assistance, where Americans risk exposing sensitive data to AI helpers; the open-weight model is now available on Hugging Face for third-party developers.

National Design Studio launches Ramparts

The National Design Studio, a US government technology organization, has launched Ramparts, an open-weight PII redaction model that runs entirely in the browser with no server calls required. At under 15MB, it is purpose-built to run on legacy devices, including older Android and iOS hardware, where competing models — the smallest of which weigh at least 50MB — cannot.

Edward Coristine, head of engineering at the studio, frames the core problem simply: when PII removal happens server-side, users have to trust that the server is actually doing what it claims. Ramparts eliminates that dependency by keeping personal data on-device by default.

We're launching Ramparts. It's a local first privacy model that puts people back in control of the data that they share with AI. None of the frontier models will actually fit in a browser, so you cannot do PII removal in the browser — you have to trust that the server is actually removing the information. So we asked: what if it's just all on device? Personal data never had to leave your device. It's by far the smallest model — the other ones are at least 50 megabytes. This is 15.

Building the model

The team evaluated 72 base models before settling on a fine-tuned MiniLM architecture. They initially tried to quantize and prune OpenAI's privacy filter but could not get it small enough to run intelligently on legacy hardware. The final model is technically a fine-tune of MiniLM, trained to identify specific PII categories — phone numbers, names, surnames, and others — and then hand control back to whatever application is consuming the library. Redaction behavior is configurable at compile time, so developers decide what gets stripped and what stays.

Primary use case

The motivating scenario is government forms, not financial data transfer. Coristine points out that Americans routinely need help understanding and filling out government paperwork, and that asking an AI assistant to help with a document risks exposing PII unnecessarily. Ramparts lets that kind of query happen locally without sensitive data leaving the device. The model is available now on Hugging Face, and the studio is explicitly encouraging third-party developers to build on top of it, including browser extensions and chatbot integrations.

Studio context

Ramparts is a side project relative to the studio's main work. TrumpRX, its prescription drug product, has reached 15 million users and claims to have saved more than $500 million in drug costs. Coristine describes the studio's operating model as small teams, with two people — one designer, one engineer — as the preferred unit for a focused project, supplemented by short-term borrowing from colleagues across the team. The studio runs seven days a week, and Coristine acknowledges the talent pitch is a self-selecting one: the scale of government problems and the speed at which a small team can move are the draw, not compensation or equity.

Every deal, every interview. 5 minutes.

TBPN Digest delivers summaries of the latest fundraises, interviews and tech news from TBPN, every weekday.