Smarties
Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

Chithi explains official Indian letters (electricity bills, bank reminders, tax demands) in Bengali, Hindi or English, and checks every date and rupee amount against the letter before the reader sees it. These letters come in formal English or official Hindi, and many people who get them can’t read that. An AI explanation helps, but a wrong due date or amount in it is worse than none. You photograph the letter. Chithi shows the answer first: how much to pay, the last date and how many days are left, and what happens if you don’t. Tap any point to see the piece of your own photo where the letter says it. You can listen to it in Bengali or Hindi, add a calendar reminder, share it on WhatsApp, or ask a follow-up question by voice. It’s built as a verifiable agent. The model makes claims, each claim quotes the letter, and a check with no model in it decides what is shown. Every date and amount in an explanation must be in the quoted words and in a second, separate reading of the photo. Anything that fails is shown apart as “check this yourself”, never as fact. A test button plants a wrong amount, so anyone can watch the check catch it. Measured on synthetic letters: 28 explanations, 16 of them from folded, shadowed photos, 114 dates and amounts shown, none wrong. A stand-in explainer that gets a quarter of its values wrong had all 53 planted errors held back. About 20 seconds and under five cents a letter.