Verifying scheme beneficiaries with document AI

When a scheme opens, applications arrive in thousands. Each one carries a community or income certificate, a bank passbook page, and often a mark sheet or ration card, and someone in your office checks every file by hand. Most of that time goes on simple questions: is the name the same everywhere, can the page be read, has this account been used before?

Signs you need this

• Staff are pulled from other work every time an application window opens, and the backlog still grows.
• Applications go back weeks later for a blurred upload that could have been caught on the first day.
• Duplicates surface only after payment, for example two applications with the same bank account.

What document AI can do

• Read certificates and passbooks in Tamil and English, from scans and phone photos, and pull out fields such as name, father's name, certificate number, annual income and account number. Reading text from an image is called OCR.
• Match those fields across the documents in one application. "K. Murugan" on the passbook and "Murugan K" on the certificate is a likely match. A different father's name is a mismatch to flag.
• Look across the whole batch for a certificate number, account number or photo that appears in more than one application.
• Flag pages that are blurred or cut off, so staff can ask for a fresh upload while the window is still open.
• Where a document carries a digitally signed QR code, as Aadhaar letters and e-Aadhaar do, record whether that code was checked, not just the printed text.

Every flag reaches a person with its reason shown next to the document. The system sorts the queue. It does not clear or reject anyone.

What must stay with a person

• The decision. Approving or rejecting an application is an officer's call, recorded under their name.
• Appeals. An applicant who is turned down should get a reason in plain words and a way to ask a person to look again. "The system flagged it" is not a reason.

Keep an audit log that records, for every file, what was read, what was flagged, who looked at it, what they decided and when. When an appeal or an audit query comes, this log is your answer.

Data protection and questions for any vendor

These files hold caste, income and bank details. Under the DPDP Act, your office stays responsible for personal data even when a vendor processes it, must keep reasonable security safeguards, and must make sure data used for a decision about a person is complete, accurate and consistent. Ask any vendor:

• Do the documents stay inside our office network?
• Is our applicants' data used to train anything outside our own setup?
• How well does it read Tamil certificates and handwritten entries, tested on our sample files?
• What does it do when it is unsure? It should say so, not guess.
• Can we export the full audit log, and how are copies deleted when our retention rules say so?

How we can help

We set up these checks with Vizhi, our vision and document model, Sol for Tamil and English text, and Thedal for matching across files. They run on Kaappu, our private AI that works inside your office, so applicant documents stay on your network. Officers see each flag with its reason and make every decision. If an application window is coming up, talk to us and we will go through how this fits your process.