Six uses that stuck after 60 days

The post does not say. He lists the six uses he is confident about and states he is deliberately sharing only those, leaving everything else to a linked write-up. The failures exist somewhere, but they are not in the thread - read the article if you want the negative results.

Personal. Fact-checking his own scripts, ranking his own feed, locating text inside PDFs, checking citations, grouping his own review notes and scoring agent traces are all self-directed work with no user-facing SLA. Useful as a low-risk first use, not as evidence for a customer-facing decision system.

Only that the uses were cheap enough to keep. It is a retention filter he set for himself, explicitly keeping what he expects to still be running later and discarding bigger bets. No accuracy, latency or cost measurement is attached to any of the six.

Each is a bounded classification or ranking over material you already have, where a typed judgement beats prose. That matches his separate framing of the problem: LLMs are chat models we work around by asking them to emit JSON, which can be malformed and must then be parsed and validated.

Isaac Flath, whose profile lists Documents in AI Products and who writes at isaacflath.com, where the longer version of this list, Six things I tried with Jev, is published. The post is field notes from his own work rather than a vendor or consultant pitch.

That LLMs are made for chat and are only a workaround for software logic. His image is asking a general-purpose model to write JSON for you, which can come back malformed; a decision model replaces that workaround with typed answers, which is why he calls it a big deal.

A reply from Josh Brown points to perch, described as semantic code linting with decision models. It is the same idea as Isaac's eval-over-traces item, using a decision model as a fast labelled check inside a developer workflow, though the link is offered without results.