RealityLoop

• 

  Initial AI belief with explicit assumptions and active uncertainties before human feedback.

• 

  Human correction reduces uncertainty, leading to a higher confidence score in the revised belief.

• 

  New human challenge introduces uncertainty, causing the AI’s confidence score to decrease.

Inspiration

Most AI systems present answers with strong confidence, even when uncertainty exists. When humans correct AI systems, users rarely see how the AI updates its reasoning or whether confidence should change at all. This lack of transparency makes it difficult to trust AI outputs, especially in sensitive or ambiguous situations.

What it does

RealityLoop is a human-in-the-loop AI system that makes AI beliefs, uncertainty, and confidence evolution visible. Users ask a question and receive an initial AI belief that includes an answer, a confidence score, assumptions, and explicit uncertainties. Users can then challenge the AI with human feedback, after which the AI revises its belief, explains what changed, and updates its confidence. Confidence may increase, decrease, or remain the same depending on how uncertainty changes.

How I built it

RealityLoop is built using Next.js and TypeScript, with shadcn/ui and Tailwind CSS for the interface.
The AI logic is implemented using Google Genkit, which orchestrates calls to the Gemini 3 API for belief generation, uncertainty modeling, and belief revision.
The application is deployed on Vercel and runs entirely online without authentication.

Gemini 3 & Genkit integration

RealityLoop uses Google Genkit as the AI framework to structure prompts, manage belief states, and handle iterative human feedback loops.
Through Genkit, the app integrates the Gemini 3 API (Flash variant) to:

• Generate initial AI beliefs
  
• Explicitly list assumptions and uncertainties
  
• Revise beliefs after human challenges
  
• Dynamically adjust confidence scores based on new information
  

Gemini 3’s fast reasoning enables real-time interaction, making confidence evolution observable and explainable.

Challenges

One of the main challenges was ensuring that AI confidence changed meaningfully rather than always increasing. Designing prompts that allowed confidence to decrease when new uncertainty was introduced required careful iteration. Another challenge was managing API usage efficiently while maintaining a smooth and responsive user experience.

What I learned

This project demonstrated how Genkit + Gemini 3 can be used not just for answer generation, but for building transparent, uncertainty-aware AI systems where human feedback meaningfully influences reasoning and confidence.

What’s next

Future work includes visualizing belief history over time, adding deeper uncertainty metrics, and applying RealityLoop to domains such as healthcare, education, and policy analysis.

Built With

• gemini-3-api
• genkit
• next.js
• shadcn/ui
• tailwind-css
• typescript
• vercel