On this page
← ML works

PrivEx

1 min read · Multimodal AI · Privacy · LLMs · Systems

problem
Sensitive information hides in documents, images, audio and video, and often only contextually.
approach
A local engine with bundled OCR, face and speech models, a desktop app with human-in-the-loop review for images, and a Chrome extension that warns before personal data is pasted into a website. No LLM and no cloud calls.
stack
Multimodal AI · Privacy · LLMs · Systems
links
Code ↗

What it is#

PRIVEX finds and redacts personal data in text, images, audio and video, entirely on your computer. It looks for names, phone numbers, emails, card, bank and ID numbers, dates of birth, addresses, faces and spoken personal details. It ships as a desktop app and as a Chrome extension that warns you before you paste personal data into a website.

Why it exists#

Sensitive information hides in many kinds of files, and often only through context: a number is a card number because of the word next to it. Most tools that handle this send your data to a cloud service or an LLM. PRIVEX is built so that nothing leaves the machine: no LLM, no cloud calls and no GPU needed.

How it works#

For text, each recognizer produces candidates with a score. Checksums raise or lower confidence (Luhn for cards, Verhoeff for Aadhaar, mod-97 for IBAN), and nearby context words such as "DOB" or "account no" push the score up or down. A small spaCy model handles names and places. Overlaps are resolved by score and then by span length, and every finding carries a human-readable reason.

For images, OCR lines are arranged into reading order and each finding is mapped back to a box that you review before it is applied. For audio, a local speech model gives word times so the matching words can be muted. For video, frames are sampled for text and checked for faces, and moving text is kept covered between samples.

Stack#

Python, FastAPI bound to 127.0.0.1, spaCy, RapidOCR, YuNet face detection, faster-whisper, ffmpeg, a web UI and a Chrome extension.

Status#

The repository reports a held-out F1 of 0.927 on 3,000 documents from the Nemotron-PII benchmark, and 0.860 on the TAB set of real court cases, and it lists the weak spots openly: names with non-ASCII letters, some place and court names, and a few ID formats. The Windows build is the packaged one.