Aegis - AI guarding the story you didn’t mean to tell

Inspiration

Our initial focus was on women’s safety. As a team of women, our discussions evolved from physical security to the issue of stalking — and then to the realization of how often stalking now begins online.

We recognized that social media has normalized a high frequency of sharing — locations, routines, background details in photos, captions, and engagement patterns. While each of these signals might seem harmless on its own, aggregated over time they create a detailed behavioral map.

Our primary concern became this cumulative exposure, and that is the core problem we set out to address.

What It Does

We are building a B2B Digital Exposure Intelligence layer that integrates directly into platforms like Instagram.

Every user generates a digital trail — posts, captions, tagged locations, comments, and engagement behavior. Our system analyzes these signals within the platform ecosystem to identify exposure patterns that may increase vulnerability.

The System: Detects cumulative exposure signals Generates a dynamic risk score based on historical + new posts Forecasts vulnerability trends Provides real-time posting guidance through an AI companion

Activates protective guidance and an evidence vault if high-threat patterns are detected

Because it is embedded infrastructure sold to platforms (e.g., Meta), user data never leaves the ecosystem. We do not create a new data-sharing surface.

How We Built It

We initially considered building a standalone consumer app. That approach created privacy and security contradictions: we would be asking users to centralize even more sensitive data in a new system.

We pivoted to a B2B service model to align with responsible AI and data minimization principles.

Stack & Tools: Figma Make + Lovable for rapid prototyping GitHub Copilot for model scaffolding Claude and ChatGPT for structured research and threat modeling

Hugging Face dataset (~31,000 image-caption pairs) as a base reference dataset

Challenges We Ran Into

1. Model Framing Exposure risk is contextual and cumulative. There are no standardized labels for “digital vulnerability,” so we had to define our own scoring logic.
2. Dataset Limitations Public datasets do not model behavioral history over time — which our system requires.
3. Prompt Iteration Achieving consistent and accurate results from AI tools required extensive prompt refinement.
4. Prototype Constraints Our prototyping tools did not support real-time collaboration, which slowed parallel work.
5. UX Sensitivity Integrating this into the user’s current flow without disruption required careful design. ----------------------------------- ## Accomplishments
6. 78% Model Accuracy Despite the absence of standardized exposure-risk labels and the need to simulate longitudinal posting behavior, we built and evaluated a probabilistic model that reached 78% accuracy within the hackathon timeframe.
7. Seamless Flow Integration We designed the intervention layer to sit directly within the posting journey — without forcing users into a separate dashboard or audit tool. Risk feedback appears contextually at the moment of posting, minimizing friction and preserving platform-native behavior.
8. Built in 48 Hours Within a two-day hackathon, we reframed the problem, pivoted from a consumer app to a B2B infrastructure model, structured a custom dataset approach, developed a working prototype, and demonstrated risk scoring with live flow integration. ----------------------------------- ## What We Learned Risk modeling for behavioral exposure requires custom labeling logic. Longitudinal data simulation is critical for meaningful vulnerability forecasting. AI-assisted prototyping dramatically accelerates iteration — but prompt precision determines output quality. Sensitive safety interventions must balance

Built With

• claude
• copilot
• figma
• lovable