PrivyLens

🌱Inspiration

We kept running into the privacy paradox: people care deeply about privacy but still share photos that leak IDs, bank cards, license plates, MRZ lines, and more. Manual redaction is slow and error-prone, so our idea was simple: make protection the default—detect sensitive content automatically and blur it before an image ever leaves the device.

Problem Statement

Personal images often contain sensitive identifiers—passports, national IDs, driver’s licenses, bank/credit cards, medical documents—that can be exposed when photos are backed up, shared, or displayed on-device. Manual redaction is error-prone and rarely done.

🛠️ What We Built

PrivyLens is a privacy-first redaction agent that:

Finds card/ID-shaped objects in photos,

Extracts and validates numbers inside them (PAN, MRZ, national IDs),

Blurs exactly those regions (pixel-accurate), and

Optionally hands ambiguous cases to an LLM (e.g., Qwen-VL) for a second look.

Minimal flow we follow (from the slide):

Image Upload (Dash)

Model Setup (pick device, load Qwen2.5-VL)

Inference (zero-shot detections + OCR text)

Obfuscation (blur/mask via classic CV)

Output (preview + save), with optional re-run on the blurred image to verify nothing leaks.

Tech

Python, Qwen

Built With

• maestro
• python
• qwen
• supervision