Inspiration
What it does
How we built it
Inspiration The project was inspired by the critical need to transform passive job-site surveillance into active, actionable safety intelligence. While construction sites generate hours of video, manual review is often inefficient and prone to human oversight. My goal was to build a system that identifies hazards automatically, ensuring a safer work environment.
How I Built It The system is built on a cloud-native AWS Bedrock architecture:
Storage: Amazon S3 serves as the central hub for raw .mp4 footage and processing outputs.
Data Sources: I utilized diverse datasets sourced from YouTube, Kaggle, and Hugging Face to ensure the model could handle various job-site conditions.
Indexing: I implemented Marengo Embed 3.0 to perform asynchronous video embedding, creating a searchable clip_index.json that maps visual segments to vectors.
Reasoning: Pegasus 1.2 provides the "brain" for the system, performing video-to-text reasoning to detect PPE non-compliance and hazardous proximity to machinery.
Interface: A Streamlit dashboard was developed to allow safety officers to run scans and export structured JSON or CSV reports.
Challenges we ran into
Accomplishments that we're proud of
What we learned
What's next for GEO4
Built With
- 1.2:
- ai
- amazon-web-services
- answer
- apis:
- bedrock
- capabilities
- contextual
- employed
- its
- languages:-primarily-python-(97.3%)
- leveraging
- model
- pegasus
- perform
- reasoning
- safety
- scans
- specifically
- to
- video-to-text
Log in or sign up for Devpost to join the conversation.