RuntimeTerrors_AISlopGuard - About Project

Inspiration

We built RuntimeTerrors_AISlopGuard because social feeds are getting harder to trust. AI-generated videos and images now look very real, and most users do not have a fast way to judge authenticity.
Our goal was to build something practical: a tool that gives quick detection, explains why it made the decision, and improves over time based on user feedback.

What it does

RuntimeTerrors_AISlopGuard scans a social media content URL and predicts whether the content is:

likely_ai
unclear
likely_human

It combines platform signals, community signals, and model inference, then shows:

final verdict
confidence band
evidence messages
raw model score vs personalized model score

It also supports personalization:

global bias (applies to all future scans)
creator bias (applies to all future scans from that creator)

So the system becomes better aligned with the user's correction patterns over time.

How we built it

We built a full-stack system:

Mobile app: React Native + Expo
Backend API: FastAPI
Model inference service: Python service
State management: Zustand
Persistence: AsyncStorage
Data layer: React Query + fetch

The final score is computed as:

$$ S = 0.30P + 0.50C + 0.20M_p $$

Where:

(S) = final score
(P) = platform score
(C) = community score
(M_p) = personalized model score

Personalized model score:

$$ M_p = \mathrm{clip}\left(M_{\text{raw}} + b_g + b_c,\ 0,\ 1\right) $$

Where:

(M_{\text{raw}}) = raw model score
(b_g) = global bias
(b_c) = creator bias

Verdict thresholds:

$$ \text{Conservative mode: } \begin{cases} \text{likely_ai}, & S \ge 0.70 \ \text{unclear}, & 0.50 \le S < 0.70 \ \text{likely_human}, & S < 0.50 \end{cases} $$

$$ \text{Non-conservative mode: } \begin{cases} \text{likely_ai}, & S \ge 0.75 \ \text{unclear}, & 0.55 \le S < 0.75 \ \text{likely_human}, & S < 0.55 \end{cases} $$

Bias updates for corrective feedback:

$$ b_g \leftarrow b_g + 0.02d $$

$$ b_c \leftarrow b_c + 0.10d $$

Direction term:

$$ d = \begin{cases} +1, & \text{if user votes AI and model predicted Human} \ -1, & \text{if user votes Human and model predicted AI} \ 0, & \text{otherwise} \end{cases} $$

Challenges we ran into

Keeping backend verdict logic and mobile personalization verdict logic perfectly aligned.
Separating personalization voting from community voting so they do not interfere.
Handling creator-bias edge cases (same creator, different content, app restarts, deletion/reset).
Building transparent UI for explainability without making the interface noisy.
Handling local dev network issues (LAN IP changes causing mobile request failures).

Accomplishments that we're proud of

Built a complete working product, not just a model notebook demo.
Added explainable evidence and score transparency instead of a black-box result.
Implemented creator-level and global personalization that persists and evolves with feedback.
Added creator-bias management tools and rich scan history details in the app.
Stabilized tricky edge cases in bias updates and feedback flows.

What we learned

Explainability is essential for trust in real-world AI products.
Multi-signal scoring is more robust than relying only on model output.
Product reliability depends heavily on state design and UX flow, not only model quality.
Small threshold mismatches across services can cause major behavior inconsistencies.
Iterative user testing is critical for making AI-assisted decisions usable and fair.

What's next for RuntimeTerrors_AISlopGuard

Add optional cloud sync for personalization across devices.
Improve creator identity resolution across platforms.
Add stronger abuse/noise resistance for user feedback inputs.
Expand model evaluation with broader real-world datasets.
Build lightweight moderation analytics for teams and creators.
Improve calibration and confidence reporting for more precise decision boundaries.