I think instead of "performance" you must mean "accuracy". Traditional deterministic conversion will always be faster and cheaper than running through a model, even if it is less accurate.
I tried to make exactly this a year ago. Built on Cloudflare using all of their primitives: https://crawlspace.dev -- It didn't work too well (so don't bother trying it).
I have no plans in downloading Atlas either, but I think your browsing isn't used for training unless you opt in.
> By default, we don’t use the content you browse to train our models. If you choose to opt-in this content, you can enable “include web browsing” in your data controls settings. Note, even if you opt into training, webpages that opt out of GPTBot, will not be trained on.
I am so sick of these anthropomorphized names that have nothing to do with anything that we’re all supposed to remember now. Why are we giving products first names? The worst offender is probably Amazon Rufus. It’s all so dumb and I hate it. At least attempt to be clever and name it something that relates to the product itself. Even Google Wave, despite its shortcomings, made sense as a product name.