Inspiration
Our project started from a simple observation: public transit safety is often reactive rather than proactive. From our own experiences on buses and trains, we noticed that incidents—ranging from harassment to suspicious activity—sometimes went unreported or were addressed too slowly. When we looked into the data, the numbers were staggering. Studies indicate that public transport crime may be 25–30 times higher than reported, and in Canada alone, nearly 4,000 sexual assaults or harassment incidents were recorded across the largest transit systems between 2013 and 2017, 90% targeting women.
Yet, existing surveillance systems are outdated. Cameras exist, but they often suffer from operator fatigue, insufficient staffing, and inconsistent reporting. Many transit agencies simply lack the capacity to monitor and act on real-time incidents. This combination of systemic gaps and personal experience inspired us to create a tool that could enhance operator awareness and response.
What it does
Aegis is an AI-powered transit monitoring app that gives operators real-time situational awareness. It ingests multiple camera feeds, evaluates the danger level of each feed using AI, and prioritizes the most critical streams. The system also provides clear explanations for why a feed was flagged, helping operators respond faster and more confidently to incidents.
How we built it
We built Aegis with a Vite + React frontend deployed on EC2, and a FastAPI backend also running on AWS EC2. Camera feeds are streamed in real time using WebSockets, and the AI processing for threat detection is powered by AWS Bedrock.
Our AI models analyze object interactions, detect anomalies, and assign a danger score to each feed. Multithreading ensures that multiple streams are processed simultaneously without lag, and the frontend dynamically displays the highest-priority feeds with concise, actionable insights.
Challenges we ran into
WebSocket streaming: Handling multiple high-bandwidth camera feeds simultaneously without dropped frames or lag.
Model validation: Ensuring AI outputs were reliable and low-latency
Integration: Combining frontend, backend, and live AI processing into a cohesive system required careful multithreading and concurrency handling.
System performance: Keeping real-time processing smooth while running AI inference on multiple streams was a constant optimization challenge.
Accomplishments that we're proud of
Successfully integrated real-time AI analysis with live video feeds.
Built a responsive, operator-friendly interface that highlights the highest-risk feeds.
Created a system that demonstrates how AI can enhance public safety without replacing human decision-making.
What we learned
Real-world data is often incomplete and under-reported, which makes model validation challenging.
Concurrency and multithreading are critical when processing multiple streams in real time.
Integrating AI into an operational system requires careful attention to both user experience and technical performance.
What's next for Aegis
Expand AI capabilities to detect more nuanced threats, including harassment and suspicious behavior patterns.
Integrate with incident reporting systems to create a full feedback loop for operators.
Explore edge AI deployment to reduce latency and allow processing directly on transit vehicles.
Conduct user studies with transit operators to refine the UI and threat prioritization.
https://github.com/David-Rodriguez-Barrios/HackTheChange2025
Built With
- amazon-ec2
- amazon-web-services
- aws-bedrock
- fastapi
- javascript
- python
- react
- vite
Log in or sign up for Devpost to join the conversation.