Josh Pigford lists where Jev runs inside two live apps
The pattern is cheap verification calls wrapped around expensive models, not a model swap.
Josh Pigford posted a long list of where he is using Jev in production across two of his apps, Granite, a document vault, and KeptWell, a family medical binder. It is dozens of separate jobs, and almost none of them are the flashy agent demos that have been circulating.
Lots of cool experimental Jev (@typesafeai) stuff getting posted lately, but what about using it in an existing product?
Detect emergency or distress before the model streams a token
In Granite, Jev second-opinions Gemini's document classification and flags low-confidence results for review, scores the quality of a PDF text layer and routes bad ones to OCR, verifies each extracted field against the page text, works out whether a document asks you to pay, sign, renew or respond, and judges whether two near-duplicate files are the same document or a revised version. The Ask feature uses it to route a question to the right tool path instead of a regex, then checks the final answer is supported by the cited passages and hedges if it is not.
The checks sit around the big models, not instead of them
KeptWell is where the pattern gets clearest. Pigford says Jev verifies every extracted lab value, dose, diagnosis and provider against the source page, flags anything it cannot confirm with a quiet marker, and catches diagnoses stated more precisely than the page actually says. In chat it detects emergency or distress before the model streams a token, reranks retrieved chunks against the question, picks between two contradicting family facts, and filters false positives out of the PHI audit, his example being the word "Ray" inside "x-ray".
Then there is a section he labels cost gates. Jev decides which lab trends are worth an Insight before calling Opus, skips the highlight call when nothing is worth highlighting, skips reprocessing documents a prompt change would not affect, routes extraction to batch or sync by urgency, and tells a bulk import apart from a runaway loop at the spend cap. He also uses it to flag uploads containing instructions aimed at an AI.
The reusable idea here is not the model. It is that a cheap call can decide whether the expensive call happens at all, and can check the expensive call's output afterwards. Everything on Pigford's list is a small, boring judgment that used to be a regex, a lookup table or a human, and most of them sit either side of a bigger model rather than replacing it.
He framed the post as a deliberate counterweight to the experiments filling timelines, asking what using this stuff inside an existing product actually looks like. Pigford has been running Jev at volume for a while, including earlier when he pushed over 10,000 Hacker News comments through it. These are his own descriptions of his own systems, with no cost or accuracy numbers attached this time.