I'm the person you find if you take the intersection of the sets of experts in accessibility and knowledgeable people about modern Transformer architectures and I feel like I fundamentally do not understand how you can reasonably claim to be doing what you're doing.

Heuristic-based automatic remediation is one, not so great, thing. But trying to probabilistically remediate accessibility based on ... The huge corpus of inaccessible web content out there as a baseline? Or possibly a far-too-small annotated training set that can capture a tiny portion of what can go wrong with accessibility where it would make more sense to just build the darn heuristics... I must be missing something here.

Also you might want to call this a Show HN.