Somewhere in India right now, someone is typing a PAN number from a scanned card into a loan form. Then the GSTIN from a registration certificate. Then the sanctioned amount from a bank letter, digit by digit, double-checking the lakhs and crores.
We kept running into this. A loan or subsidy file is a stack of twenty or thirty documents, and almost all the work is reading them and copying the right values into the right boxes. It's slow, it's error-prone, and nobody enjoys it.
So we built patrapi. Patra (पत्र) is a letter or document; the rest is what you get back.
document in, api out
The idea is simple:
- Upload a sample of the document.
- Say which fields you need, in plain words. "The sanctioned amount, the interest rate, the borrower's name."
- Test it on a few real files.
- Publish, and you get an endpoint your software can send documents to.
There's no training run and no annotation. If you can explain a field to a new intern, you can build the workflow. And if you're not sure what to ask for, the assistant reads your sample, suggests fields with types and rules, and tests them before you publish anything.
every field shows its work
The part we cared about most is trust. A number on a loan file has to be right, and "the AI said so" isn't good enough.
So every field comes back with the value as printed, a cleaned-up version ("₹ 48,50,000/-" becomes 4850000), the page it was on and the exact text it was read from. When something isn't on the document, patrapi says so instead of making something up. Anything uncertain can wait in a review queue for a person to check.
starting with india's paperwork
patrapi works on any document, but nobody should start from a blank page. So we wrote recipes for the paperwork behind MSME loans and machinery subsidies: PAN, Aadhaar, GST REG-06, Udyam, ITR, sanction letters, Index II, electricity bills and a few dozen more. Use a recipe as it is, or fork it and add the field your bank asks for.
You can browse them all at patrapi.com/recipes.
what happens to the documents
These are some of the most sensitive documents people own, so we were strict about this. Documents are encrypted while a job runs and deleted once it's done. If a result needs a person to review it, the document waits until it's reviewed, and never more than seven days. They are never written to our logs and never used to train anything. Your clients' documents are not our training data.
try it
You get 50 free credits when you sign up, no card needed. After that it's from ₹5 a document.
If you spend your days with loan files, subsidy applications or KYC paperwork, I'd love to hear what you're copying by hand. Send me a sample of the document, and I'll build the workflow with you.