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 Roll Call Vision
What it does
Volunteer boards approve minutes that record each vote; transcripts often get the count wrong. Roll Call Vision uses OpenCV 5.0 to watch the meeting video during each vote window, count raised hands per seat, and compare with the transcript. An agent then confirms the count, re-looks at the window more densely, or holds minutes approval and asks a human (fail-closed).
Results
Synthetic eval: 32 meetings / 81 votes, 0 unsafe confirms. Count accuracy is measured on our own synthetic clips; failure cases (occlusion, half-raised hands) are listed in REPORT.md.
Live on AWS
Deployed on AWS Lambda (us-east-2) as a zip function plus an OpenCV 5 layer, behind a Function URL: https://hj6wfy2lilbbllikwriblbsx3m0thfif.lambda-url.us-east-2.on.aws/
The Function URL uses AWS_IAM auth, so unsigned requests get a 403. The demo video shows live calls. Message us through Devpost and we will run all four demos live over a screen share at a time that suits you. Everything else runs locally in a few minutes: the README has the test suite, the four-case demo and the 32-meeting evaluation.
Code, technical report and architecture: https://github.com/by42ppcrps-dev/roll-call-vision (REPORT.md, docs/ARCHITECTURE.md). REPORT.md includes an agent trace where the first OpenCV pass changes the next step.
Built with
OpenCV 5, Python, AWS Lambda and CloudWatch. Planned, not deployed: an S3 video input and a Board Clerk MCP tool that calls this service.
Built With
- amazon-web-services
- aws-lambda
- cloudwatch
- opencv
- python
