What is Jev?
If you've used an AI recently, like Claude or ChatGPT or any traditional LLM via an API, you're paying a big price for high latency and things your software doesn't need. LLMs are made for humans, but when you try to use them inside of your backend infrastructure, for automation, or parsing, they're really slow & expensive.
TypeSafe AI built Jev, a System One model. Instead of writing paragraphs, Jev takes your data in and gives you a clean, typed true/false answers or categories in about 80 milliseconds.
How I used it
I wanted to build an automated workflow that processes my inbox, extracts my e-receipts, and logs them into a budgeting database.
Using traditional LLMs to parse every single incoming message was not an option; waiting 8 seconds per email completely ruins throughput, and the API bills skyrocket. Basic regex scripts failed because every merchant format (Amazon, Uber, local coffee shops) structured their receipt bodies differently.
Where Jev comes in:
- A webhook triggers whenever a new email lands in the inbox.
- The raw email text gets passed to Jev. It handles two specific schemas in parallel: a Choice schema for email categorization
(is_receipt, is_newsletter, is_personal)and a Noul boolean schema(requires_urgent_action). - If it's a receipt, a second fast pass extracts a structured Choice category for the expense type
(Software_SaaS, Travel, Food, Hardware)alongside a Score query for the confidence interval. - Jev returns these type-safe variables in 80-114 milliseconds.
Because Jev has a 0% schema format error rate, I didn't need to write any text-parsing error handlers or try/catch wrappers. If the confidence score is high (> 0.95), the line item writes straight to the DB table.
Why this makes sense
It is made for high throughput data automation:
- Parsing Receipts: Turns transaction emails into structured DB rows instantly.
- Sorting Inboxes automatically, tags and routes support tickets, webhooks, or emails instantly.
- Filters broken or useless data.
It costs just $42 per billion tokens. That makes it roughly 238x cheaper than trying to build this same pipeline with a big chat model.