General Experiment Visualizer

Inspiration

During my physics studies, I encountered several abstract and difficult-to-grasp concepts, such as the “virtual object” in geometrical optics. Even after multiple discussions with teachers and classmates, I only half understood it and often made mistakes when solving problems. It wasn’t until I watched an experiment video and drew the ray diagram that I truly understood and absorbed the concept.
This experience made me realize how important a tangible, visual experiment or demo is for internalizing abstract concepts into one’s own knowledge.
At that time, I was self-learning Swift, so I used the double-slit interference experiment in geometrical optics as a practice project. I was pleased with the result, but there are far too many physics experiments to code them one by one, so I came up with the idea of using a large language model to generate these demos automatically.

What it does

This project is an interactive experiment generation agent powered by Perplexity MCP (AI search) and GPT-5.
When a user types a description of an experiment they want to generate into the chat box on the homepage, the agent will:

Interpret the user’s request and determine the target experiment.
Use Perplexity MCP to retrieve the experiment’s process and adjustable parameters.
Have GPT-5 write an HTML script to display the experiment in an interactive demo with a dark theme and low-saturation colors.

How we built it

The overall framework was built by SOLO Builder after I described my idea.
The core functionality was straightforward: call the GPT-5 API and instruct the model on how to use Perplexity MCP to retrieve experiment data.
I refined the system prompt multiple times to fix the output format so the generated results could run directly.
The visual design went through multiple iterations using the select element feature, allowing me to adjust front-end code and layout intuitively, similar to using professional design tools. The final style is a dark, sci-fi aesthetic.

Challenges we ran into

Initially, GPT-5’s output quality was worse than GPT-5-mini, requiring careful prompt tuning to achieve stable results.
The first version’s three-column layout (chat on the left, code in the middle, preview on the right) was cramped, poorly colored, and looked like a scam website.
Achieving the right balance between aesthetics and usability required multiple UI and styling adjustments.

Accomplishments that we're proud of

Built a complete pipeline that automatically generates experiment demos from user input.
Delivered a final interface that matched my vision, including animated light beams with an ocean-like flowing effect.
Produced a virtual object demo that adheres fully to physical principles.
Kept the cost low (around 0.02 USD per generation) and provided public access through my own API.

What we learned

A visual, interactive experiment is incredibly effective for understanding abstract physics concepts.
System prompt engineering is crucial when using large models to generate both code and UI.
Attention to detail in layout and color design greatly enhances user experience.

What's next for Experiment-Generate Agent

Support experiment generation for more subjects beyond physics.
Add richer visual effects and interactive features to make experiments more immersive.
Improve generation speed and stability while further reducing costs.
Introduce a community feature for sharing and improving generated demos.