RealityCheck

Private, On-Device AI for Healthier Digital Wellbeing

Comment

RealityCheck

Inspiration

Social media has become an important part of how young people communicate, learn, build communities, and discover opportunities. However, it can also create constant pressure through social comparison, engagement metrics, unrealistic lifestyles, emotionally harmful interactions, and fear of missing out.

This inspired us to ask:

Can AI help people build healthier relationships with social media without sending their personal content to a cloud service?

That question led to RealityCheck, a privacy-first, on-device AI digital wellbeing system.

What We Built

RealityCheck is designed to detect digital wellbeing risks in online content and provide timely interventions.

The system focuses on signals such as:

  • Social comparison
  • Emotional pressure
  • Engagement pressure
  • Fear of missing out
  • Manipulative or urgency-driven language
  • Negative digital interactions

Instead of simply telling users that content is "good" or "bad", RealityCheck provides a wellbeing-oriented risk assessment and encourages healthier digital behavior.

Our architecture is designed around local inference:

Online Content ↓ Local AI Analysis ↓ Risk Detection ↓ Wellbeing Assessment ↓ Personalized Intervention

The project uses a compact GGUF language model intended to run locally through llama.cpp, allowing the core AI analysis to work without sending sensitive content to a remote AI API.

How We Built It

We built RealityCheck as an offline-first system.

The project combines a browser-based user interface with an on-device AI layer. The AI analyzes relevant text and identifies patterns associated with digital wellbeing risks.

We also designed a wellbeing scoring system that combines several signals, including screen time, detected pressure signals, interventions, AI analyses, and recovery behavior.

The objective is not to diagnose a person or replace professional mental-health support. RealityCheck is designed as a digital wellbeing and awareness tool that helps users recognize potentially unhealthy online patterns.

Privacy was a central design principle. Wherever possible, analysis is intended to happen locally on the user's device rather than requiring continuous cloud connectivity.

What We Learned

Building RealityCheck taught us that deploying AI on constrained hardware requires much more than choosing a powerful model.

We learned about:

  • GGUF model formats
  • LLM quantization
  • Local LLM inference
  • llama.cpp
  • On-device AI
  • Natural language processing
  • Browser-extension architecture
  • Privacy-preserving AI
  • AI risk classification
  • Resource-constrained deployment
  • AI benchmarking and profiling

One of our biggest lessons was that an AI system must balance accuracy, latency, memory usage, privacy, and usability.

A model that performs well in a cloud environment may not be practical on an ordinary laptop. This challenge pushed us to think about AI from an accessibility and deployment perspective.

Challenges

One of our biggest challenges was designing an AI system that could provide meaningful contextual analysis while remaining small and efficient enough for constrained hardware.

Another challenge was reducing false positives. Online language is highly contextual, and the same word or phrase can have very different meanings depending on the conversation.

We also had to consider unreliable connectivity. Rather than making continuous internet access a requirement, we designed the core AI architecture around local inference.

Finally, integrating AI analysis with a practical wellbeing experience required us to think beyond the model itself. The AI output must be understandable and useful to the person receiving it.

Why It Matters

RealityCheck explores a different approach to AI accessibility:

Useful AI should not always require a powerful cloud server.

Many people use affordable devices and may have limited or unreliable connectivity. By exploring efficient local inference, we want to demonstrate that privacy-conscious AI applications can be designed for the hardware people actually have.

Our longer-term vision is to expand RealityCheck with more African languages, more efficient models, stronger contextual understanding, and broader digital wellbeing capabilities.

Future Direction

We plan to improve RealityCheck by:

  1. Benchmarking multiple compact models.
  2. Improving contextual risk detection.
  3. Supporting more African languages.
  4. Reducing false positives.
  5. Improving intervention personalization.
  6. Expanding offline functionality.
  7. Optimizing inference for affordable laptops and devices.

Conclusion

RealityCheck started with a simple idea: technology should help people use technology more consciously.

By combining digital wellbeing principles with privacy-first, on-device AI, we are exploring how AI can become more accessible, practical, and respectful of user privacy.

RealityCheck is our step toward a future where intelligent digital wellbeing tools can work locally, privately, and efficiently — even on everyday hardware.

Built With

Share this project:

Updates