Inspiration
What it does
How we built it
Challenges we ran into
Accomplishments that we're proud of
What we learned
What's next for CaptureGate
Inspiration
A photo can look clean after cropping and still have lost an important edge or line. CaptureGate focuses on the handoff between capturing a document and sending it onward: make excluded content visible, preserve failure findings, and ask the person to approve the exact version.
What it does
Upload a fictional document photo, try automatic page detection, or select four corners manually. Compare the original and proposed image; undo cropping, rotate, and choose natural color, contrast cleanup or black-and-white. Orange crop-loss shading shows excluded areas. An accepted handoff contains reviewed.jpg, review.json with the output SHA-256, and readable warnings. Changing the proposed version clears acceptance. Human approval never converts a failed whole-page check into an automatic pass.
How we built it
OpenCV 5 performs substantive image decoding, contour and quality analysis, perspective correction and cleanup in a shared Python service. The browser provides border editing, comparison and exact-version acceptance. On AWS, Lambda runs the container stored in Amazon ECR; API Gateway exposes a password-protected judge demo, and DynamoDB atomically caps lifetime analysis attempts. The direct Lambda URL remains IAM-protected. Local operation requires no paid model or API key.
Challenges and lessons
Weak edges and nested rectangles can defeat automatic boundary detection. We added reversible four-corner correction instead of pretending every automatic crop was reliable. We also separated a human-accepted proposal from the original quality verdict. A deployment file-permission failure was caught before opening the gateway, corrected, and retested under a restricted runtime user.
Evidence and limitations
75 Python tests passed in Linux; 46 controller tests passed locally. Seven synthetic Lambda checks and live judge-access checks verified processing, credential rejection and exact UI bytes. Export contents, CRC and image hash were checked locally. Six synthetic crop cases evaluate geometry, not real-world accuracy or measured user benefit. Development samples are not held-out validation.
Cleanup cannot recover missing text, severe blur or hidden page edges, certify authenticity, or prove completeness. No OCR improvement or time-saving claim has been measured. Physical-touch evaluation and user studies remain future work. Use fictional documents for the demo; remote analysis sends the image to the service.
Team and next steps
Built by Redwan Rahman with AI-assisted implementation and documentation. Next: independent capture evaluation, touch-device testing and measured downstream usability. We enter the overall competition; no COOL or Agentic Vision special-prize qualification is claimed.
Code and technical report
Public MIT-licensed source, pinned dependencies, architecture, build/test instructions and evaluation limitations: https://github.com/redwanrahman-cloud/capturegate
Technical report: https://github.com/redwanrahman-cloud/capturegate/blob/main/TECHNICAL_REPORT.md
Judge demo: https://yq4tqjkm2f.execute-api.us-east-1.amazonaws.com — password supplied privately to judges; access expires November 11, 2026 at 00:00 UTC.
Log in or sign up for Devpost to join the conversation.