PrivacyLens
A privacy-focused desktop experiment for real-time screen redaction.
Established a modular capture -> detection -> overlay architecture that can grow into a practical privacy tool instead of a one-off computer vision demo.
- Role
- Solo Developer
- Team size
- Solo build
- Updated
- 2026-04-07
Overview
What the project is and why it mattered.
PrivacyLens explores how to analyze on-screen content in real time and visually redact sensitive areas. The project is early, but the architecture is designed to make experimentation with models, overlays, and capture sources easier over time.
Architecture
Capture -> detection -> overlay
Status
Early but functional shell
Direction
Privacy-focused CV tooling
Problem
Real-time privacy tooling has to balance responsiveness, clarity, and extensibility. I wanted to build a pipeline that could eventually support both pretrained and custom ONNX models without tightly coupling every piece together.
Role
I designed the architecture, built the app lifecycle and overlay system, and structured the pipeline so capture, detection, and rendering could evolve independently.
Build details
Stack, constraints, and decisions.
Stack
Constraints
- The system needs to stay responsive while processing live visual data.
- Future model support should not require rewriting the app shell or overlay layer.
- The project is still early, so the architecture has to support iteration without premature complexity.
Decisions made
Separate capture, detection, and overlay concerns
Breaking the pipeline into independent layers keeps the project flexible as model support and capture targets evolve.
Ship the visual system before full detection
Building the overlay window and censor rendering first made it easier to validate the UX side before full inference was wired up.
Design for custom models early
The project structure anticipates experimentation with both pretrained and user-selected models rather than hardcoding a single path.
Outcome
What came out of it.
Outcome
- Implemented the application lifecycle and overlay window system.
- Built dynamic censor-box rendering with test data to validate the UI layer.
- Set up a modular service pipeline for later screen capture and inference work.
Lessons
- Real-time systems benefit from clear boundaries between pipeline stages.
- Prototype visibility layers early so model work has a clear target.
- Privacy tooling needs thoughtful controls, not just a working detector.
Next
- Complete live screen capture integration and ONNX inference.
- Add label filtering, confidence thresholds, and multi-monitor support.
- Explore blur or pixelation options in addition to solid censor boxes.
Keep browsing
Keep moving through the archive or reach out if you want to talk through similar work.