Aegis AI

Aegis AI: Your Silent Protector

The inspiration for Aegis AI came from a simple, sobering reality: many people feel unsafe walking home alone at night, but calling for help is often impossible in the heat of a moment. We wanted to build a "Shadow" protector—a system that doesn't wait for you to press a panic button, but understands the context of your environment and acts on your behalf when you can't.

Aegis AI is an intelligent safety net. When a user enters "Shadow Mode," the app begins streaming live audio to our backend.

Real-time Understanding: It doesn't just record; it transcribes voice in real-time. Risk Assessment: A specialized AI model (Google Gemini) analyzes the conversation for markers of coercion, harassment, or imminent danger. Guardian Alerts: If a high-risk situation is detected—like a boundary being ignored or a threat being made—Aegis instantly alerts a pre-defined "Circle of Guardians" with the user's live location and the transcript of what's happening.

We built a modern, low-latency stack to ensure safety in real-time:

Frontend: A Next.js web application that handles WebSocket streaming for audio and live status updates. Backend: A FastAPI server that orchestrates the data flow. Live Transcription: We integrated AssemblyAI V3 Streaming to get near-instant text from audio. The Brain: Google Gemini (gemini-2.0-flash) serves as our risk triage engine, processing transcripts against a complex hierarchy of safety signals. Persistence & Auth: Supabase handles our user data, guardian relationships, and real-time database notifications.

The biggest technical hurdle was latency. In a safety-critical application, every second counts. Balancing the time it takes to transcribe audio, send it to an LLM, and trigger a notification required us to implement a sophisticated background assessment queue. We also faced challenges in "noisy" environments—ensuring the AI can distinguish between friendly banter and genuine distress was a major focus of our prompt engineering and system design.

We are incredibly proud of our Risk Triage Engine. Seeing the system correctly identify a "critical" risk based on subtle linguistic cues (like an ignored 'No' or an attempt to isolate the user) was a "eureka" moment for us.

Building Aegis AI taught us the power of Context-Aware AI. We learned that LLMs are remarkably good at risk assessment when given the right "grounding" (like the user's initial intent for their journey). We also deepened our understanding of WebSocket architecture and the complexities of real-time audio processing in a web environment.

Aegis AI is just beginning. Our roadmap includes:

Native Mobile Apps: For better background processing and OS-level emergency integration. Multi-Modal Detection: Adding video and movement analysis (detecting a sudden run or a fall). Offline Mode: Using local on-device models to provide basic safety checks even without a data connection. Strategic Partnerships: Working with campus safety organizations and local authorities to create a direct bridge between Aegis alerts and official responders.