In my testing I tried Gemini 2.5 Pro, Qwen 3 Coder 480B and Opus

Thanks to tools like Blender and MCP servers, models from Gemini to Qwen can be used to create fully usable models.

Having LLMs generate code to build a model by using tools like an artist would drastically increases quality over models produced directly by text-to-3d models like Hunyuan3d

More examples I was able to generate: https://imgur.com/a/3d-models-generated-by-ai-agent-qzOMpqr

I'm building an AI web agent to fetch detailed information. Looks up and browses sites just like you.

Have just over 500 developers signed up and releasing our client library soon.

https://www.onequery.app

I find v0.dev is remarkably good at this. But deploying straight from there is not frictionless atm.

Your approach makes a lot of sense. Especially for "old school" websites with simple Server Side rendered pages, tables are a treasure of data. Wikipedia tables would be a great use of your tool too.

I was thinking more on the lines of programmatic access that `enforces` structured output. LLMs are really good at this step. Define a schema and get an output that is guaranteed to fit.

You can see some examples of what I mean here: https://query-rho.vercel.app/