Marg.ai

Talent is everywhere; exposure isn't. In my village in Rajasthan, few children have anyone who asks them why, or anyone who has walked a path beyond the village. Marg is a voice-first guide on WhatsApp that brings children that exposure through the people already around them, in Hindi, Hinglish, English or Braj. It starts in rural India, and it can reach anywhere families share one phone and speak a language at home that school doesn't. Parents of 3–8-year-olds get Aaj ki Khoj, a two-minute discovery to do offline with things at home. The parent says what the child noticed, and Marg answers the child's idea with one question back. A photo of what the child collected is counted (SAM 3.1), and the discovery becomes a card (Muse Image). Children, once a parent agrees, can ask Marg their own why by voice, in short, safe sessions. Teachers get today's chapter linked to three real jobs, with a five-minute class activity. Students get a mentor that answers only from 53 official sources, cites them, and says "I don't know" when there is none. Every message in and every reply out is checked by Llama Guard 4 and Muse Spark, and nothing unchecked is ever sent. Crisis replies are fixed text with official helplines, never written by AI. Impact. Marg reaches families through the phone already in the house: almost 90% of rural 14–16-year-olds have a smartphone at home (ASER 2024), and Marg needs no app, only WhatsApp. It costs under one US cent a reply in model calls. A new region changes only three things: language word lists, official sources and crisis helplines. Built during the hackathon on six Meta models, with 3,354 tests and 101 eval cases run through the real pipeline. In unscripted field tests, 8 real people (a teacher, students and a mother) found 31 problems; each was fixed and tested the same day. Try it live: https://marg-ai.fly.dev/demo (the same pipeline as WhatsApp, with a trace of every step behind each reply).