Send it one page of a tax PDF and Jev says which IRS form it is, out of 261, and what kind of page it is, with a probability for each answer. No model training is needed.
The team at Kyoto built it to replace the language-model pipeline they used last tax season. They report it classifying their whole document set correctly while being 34 times cheaper and 6 times faster.
A TypeScript library, with the full list of forms and the rules used to recognise them.
What you get
- Full TypeScript source code, Apache 2.0 licensed
- Rules for 261 IRS forms
- PDF page reading helpers
Jev
Python
FastAPI
React
TypeScript
PostgreSQL
Rust
CodeRabbit Top10
Convex
Cloudflare Workers
Next.js
