Tripo AI 3D Generator
Mini toolsRampart
Rampart combines a neural token-classification layer (MiniLM-L6-H384 fine-tuned with a 35-label BIO head for 17 entity types) with a deterministic regex layer for structured identifiers such as SSNs, credit card numbers, email addresses, and IP addresses. The shipped artifact is 4-bit MatMul quantized with INT8 embedding (onnx/model_q4.onnx), trimmed to a 19,730-piece vocabulary, and runs at 6.6 ms p50 on CPU and 3.9 ms p50 on WebGPU. It achieves 98.42% private-term recall across 7 supported Latin-script languages (English, Spanish, French, German, Italian, Portuguese, Dutch) on a 30,000-row held-out test. Entity types include personal names, phone numbers, tax IDs, bank account numbers, government IDs, passport and license numbers, and street-line address components. Intended for redacting user-typed text before it reaches hosted LLMs, with stable placeholder substitution across multi-turn conversations.
New
Text
Gen 7
Overview
Rampart is a 14.7 MB quantized ONNX token-classification model for client-side PII detection and redaction. Fine-tuned from MiniLM-L6-H384 with a 35-label BIO head covering 17 entity types, it supports 7 Latin-script languages. Designed to run in browsers via ONNX Runtime Web and Transformers.js at 6.6 ms p50 latency.
About National Design Studio
National Design Studio is a White House Office agency redesigning how Americans experience their government, tasked with overhauling roughly 26,000-27,000 federal web portals under the "America by Design" initiative and returning time and dignity to the people it serves.
View Company Profile
Last updated: July 1, 2026
