Inspiration
Screenshots are useful evidence, but they are also a common accidental disclosure path. A support engineer can remove a password and still miss an invite QR code, customer address, face, or ticket identifier. QuietGuard Vision makes a sanitized copy and preserves the original, so privacy review becomes a repeatable gate instead of a last-second visual guess.
What it does
QuietGuard Vision receives an image locally or through an AWS HTTP API. OpenCV 5 detects QR geometry, frontal faces, and connected text regions without exporting image contents to an OCR service. A deterministic policy scores risk and chooses one of three actions: pass, create a redacted copy, or stop at a human review gate when detections cover too much of the image. It then verifies that every chosen region materially changed, writes a hash-chained audit trace, and returns the sanitized image plus machine-readable evidence.
How we built it
The perception layer uses OpenCV 5 QRCodeDetector, YuNet through FaceDetectorYN, grayscale thresholding, morphology, contours, and pixel-difference verification. The same Python package runs as a CLI and as an AWS Lambda arm64 container behind API Gateway. AWS SAM describes the deployment with a minimal execution role. A rights-safe synthetic benchmark generates varied meeting screenshots with known QR and text ground truth, reports region recall, verifies source integrity, and tests a clean control.
Challenges we ran into
Privacy tooling must avoid two bad outcomes: publishing a missed secret and destroying the only original evidence. We made the source immutable, used copy-only redaction, added a coverage-based review gate, and built verification as a separate step. We also avoided sending text to an external OCR model; the default system only identifies geometry.
Accomplishments that we're proud of
- A working OpenCV 5 pipeline with QR, face, and text-region perception.
- Perceive, decide, act, and verify orchestration with explicit human control.
- The same workload runs locally and in AWS Lambda.
- Repeatable, rights-safe evaluation with machine-readable metrics.
- Source-preservation tests and a tamper-evident audit chain.
- The current synthetic benchmark matches 40 of 40 expected regions with 1.0 precision and recall, preserves source integrity, verifies every redaction, and produces zero detections on the clean control.
What we learned
Redaction quality is only part of privacy safety. The system also needs source immutability, measurable post-action checks, bounded autonomy, and evidence a reviewer can inspect. OpenCV's classical geometry tools are useful because they can work locally without transmitting the private text itself.
What's next for QuietGuard Vision
We plan to publish AWS arm64 latency and cost measurements, add OpenCV DNN face detection as an optional stronger backend, and evaluate on consented real-world screenshots alongside the synthetic benchmark.
Responsible use and disclosure
The project was created during the competition period. OpenAI Codex was used as a coding assistant. All included image fixtures are generated synthetic data. No private screenshot or extracted text is sent to a third party by the default pipeline.
Evidence and reproducibility
- Live project walkthrough
- Public source repository
- Technical report
- Architecture and agent workflow
- Machine-readable benchmark metrics
The repository pins OpenCV 5.0.0.93 and includes the AWS SAM template, Dockerfile, tests, synthetic fixtures, limitations, failure handling, human-control policy, and generated evaluation artifacts.
Log in or sign up for Devpost to join the conversation.