Aegis

• 

  Aegis

Inspiration

We were inspired by the many brave acts of people around us to voice their experience regarding sexual acts and violence. Research shows that 1 in 5 girls and 1 in 7 boys have experienced some form of sexual violence in their life. Most of these online sexual harassment target prepubescent kids and teenagers through online messaging apps. We realized that a lot of people have experienced acts of sexual violence from lack of proper understanding between children and parents, and are repressed from voicing out due to lack of actions or evidence to report to authorities. As such, we wanted to create a safer space for children by allowing parents to do real-time monitor on any incoming messages safely without intruding their privacy 24/7

What it does

Aegis is a chat safety companion for families. The child and the parent each have their own account, connected once through a six digit pairing code. On the child's device, Aegis reads messages from chat apps the child has agreed to monitor, runs them through an Indonesian text classifier (IndoBERTweet fine-tuned for sexually explicit content) plus rule based patterns for grooming signals like photo requests, secrecy, pressure, and threats. When a message crosses the threshold, two things happen at once: the child gets a gentle popup asking if they want to report it to their parent or mark it as not a problem, and the parent gets an alert on their dashboard showing the flagged message, the confidence score, and the suspected pattern. Nothing else is stored. Safe conversations never leave the phone.

How we built it

We used fine tuned model from hugging face to classify suspicious texts and not. The F1 accuracy is about 0.81. Model that we used is a fine tuned version of IndoBERT. Then, we develop the mobile apps using python as backend and ML-handling and Flutter as the main stack that allows us to launch the app in mobile environment.

Challenges we ran into

We had varying difficulties in setting up the environment for the application especially to allow the server and backend API to connect in a mobile app environment, as this is our first time in developing a proper mobile app. There's also limited memory

Accomplishments that we're proud of

We managed to develop a real mobile app that allows real-time monitoring of online messages with a lightweight AI system.

What we learned

Writing the API contract before writing code is what let three people work in parallel without blocking each other. Testing on the real pipeline early matters more than writing more features. And for anything involving an ML classifier, rehearse the demo inputs, because the model does not care what looks obvious to you.

What's next for Aegis

• Improvements AI accuracy and models
• Improvements on monitoring system to allow more apps and branch out to other systems such as games.
• Shipping of Aegis fully as a mobile APK

Built With

• docker
• flutter
• indobert
• python