A local invoice agent that reads, decides, and flags

Surya 2 and Chandra 2 handle OCR on the laptop, while a small model decides whether to pay, file, or escalate.

A new Made by Agents project builds an invoice pipeline that runs mostly on a local machine. Surya 2, a 650-million-parameter OCR model, reads a page in about nine seconds on a Mac. Chandra 2, a 4-billion-parameter model, is slower but better on tables and handwriting. A strict schema turns the output into typed fields.

The agent layer then decides what the document is and what to do with it: pay through Stripe, file the receipt, or flag a human. No document leaves the laptop except the payment call. The author argues that extraction is no longer the bottleneck; the hard part is what happens after the text comes out.

The project also documents a failure. One agent paid a euro invoice in dollars, and the reason was not a bad OCR read but a missing constraint in the decision layer. That is why the guide recommends the smallest possible payment key, treating documents as hostile, and keeping a human approval step before money moves.