Maybe I’m misreading what you’re trying to say.

Would love to share notes. I was able to get away with landing.ai and some really careful schema design and multi-step workflow with a few agents sprinkled in at the end.

Source: me. I built it with some folks.

We use landing.ai to parse the PDF, as well as useworkflow.dev to durably perform other work such as rendering PDF pages for citations, and coordinating a few lightweight agents and deterministic checks that flag for inconsistencies, rule violations, bias, verify appraiser credentials, etc. etc. Everything is grounded in the input document so it makes it pretty fast and easy. We’re going to market soon and have an approval sign up gate currently. Plenty of new features and more rigorous checks planned to bring us to and exceed parity with competition and human reviewers.

There’s plenty of margin for cost and latency versus manual human review, which takes an hour or more and costs $100 or more.

Which brings me to... choosing to throw TypeScript under the bus for having an "unsound" type system ignores the fact that you can just throw more tooling at the problem (e.g., a linter that disallows `any`, since that one example cited in the article), and glazes over the fact that the JS ecosystem has a lot more examples in the training data for many LLMs. The further you deviate from the slop-mean, the worse time you're going to have using LLMs to generate code, slop or otherwise.

> We want AI to be aligned with all of humanity.

One of many contradictions. Liars.