UdhaarAI

The problem

Most kirana shops record customer credit (udhaar) in a paper notebook, even when they have an app, because writing is faster than typing. Apps like Khatabook and OkCredit still need every entry typed in, and generic handwriting OCR gives you text, not a ledger. UdhaarAI sits in that gap.

How it works

• At the end of the day the shopkeeper photographs the page
• The OCR reads Hindi, Telugu and English, even on mixed pages
• Each entry is checked against the database, and anything uncertain goes to a review screen with a confidence score
• One tap updates every customer's account

More than a scanner

• Action notes: type or speak a sentence in English, Hindi or Telugu, and my own parser works out the action. Anything involving money always asks for confirmation.
• Business insights like a health score, a customer trust score and the best time to send a reminder, all built as explainable maths rather than an AI guess
• Reminders written for each customer from their real history, with a checker that rejects anything threatening before it can be sent
• A no-sign-up demo with a guided tour, so anyone can try it with a sample shop

My own code first

I used plain code wherever it could do the job and kept Gemini for what it couldn't, like reading handwriting. The project has 224 automated tests across 9 test suites.

What I learned

Where AI helps and where plain code is better. It was tempting to hand everything to the model, but rules I could write and test were more reliable for money. Two days also forced me to decide what mattered most and leave the rest out.

The Knowledge Society

UdhaarAI was my entry to the TKS Prompt to Product challenge: two days to build and pitch a working AI product that helps businesses. It took me to the interview round, and I was selected for a place in the TKS virtual programme, a 10-month accelerator for teenagers that started in Silicon Valley.