Andromeda

• A LMM's output does not need to be a wall of text with millions of bullet points. Markdown is a great way to format text, but it also has limits against the sheer amount of text LLM's use to answer questions.

• The goal of this project was to use the unprecedented coding ability of modern LLM's to enhance the user experience for the normal user.

So Andromeda is a Agent that can use the LLM's ability to code beautiful and interactive Web pages to answer questions in a Very Visual and interactive way.Using the power of HTML, CSS and JS

The following are some screenshots of the output from Andromeda:

Prompt : Teach me about Orbital mechanics Visually!!

Model : DeepSeek V4 Flash 0731

As you can see Compared to using Markdown, the output is much more visual and interactive. The user can interact with the output to understand the topic better and in a more intuitive way.

Tool and Features

1. Get relevant webpages - It uses the DuckDuckGo API to search the web for relevant webpages to answer the user's question.

2. Batch read pages - It can read multiple webpages for quries per webpage. It uses jira.ai's web page reader to read the webpages and extract relevant information from them using GPT 5 Nano.

3. Search images - It can search for images using the Serp API to find relevant images to answer the user's question.

4. Update user - It can provide progress updates for tasks requiring more than 2 tool calls.

5. Send HTML response - It sends a HTML along with a guess for long for the height of the iframe container of the HTML code.

Available Models

Note : New LLM models are coming all the time so this list is volatile and will be updated as new models are released.

1. GPT-5 Nano
2. DeepSeek V4 Flash 0731
3. GPT-OSS 120B
4. Qwen 3.5 Flash
5. Nemotron 3 Super 120B
6. Gemma 4 26b
7. Inception Mercury 2

Final Notes

• There are a few things i learnt making this project.

  1. Dynamic iframe height is a pain in the ass.
  2. Modern LLM's are suprisingly good at coding beautiful and interactive Web pages. (even more than i expected)
  3. Intent Classification is tricker than it looks. (i had to use a few tricks to get it to work properly)
  4. Generative UI UX will definitely be a thing in the future. (I can see myself using it quite often)
  5. Good image API's are hard to find. Serp API is very good and very limited.

• This project is part of a series of Agentic Projects that i am working on.

• Althouth i suspect not many will, you are full free to Contribute to this project. (i will be happy to accept any PR's).

Thank you for checking out this project and i hope you find it useful.