Skip to Main Content

Ventriloquist

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

Ventriloquist — Demo video

Ventriloquist gives a voice back to awake ICU patients who can't speak because of a breathing tube or tracheostomy. A phone on the bed rail asks yes/no questions and reads whatever movement the patient can still make: raised eyebrows, a long blink, a finger lift. Each question is the one that narrows things down fastest, given the patient's own measured error rate and what is going on in the room (what the nurse said, the bedside photo, the visitor who just left). When one message reaches 70%, the phone asks "Is this your message?", and it speaks, in the first person, only after a YES lifts it to 95%. Anything not on the list gets spelled, with word completions. Three quick YES signals raise an alarm on every screen. A caregiver console shows every answer, the evidence behind each guess, and each model call with its latency and cost. A spoken message stays on that screen until someone taps "I'm coming". With the patient's YES and a caregiver's approval, an agent can send the message to a relative on WhatsApp and screen the reply before the patient hears it. It works in English and Indonesian. I'm an emergency medicine intern in Indonesia and built it solo during the hackathon. On my own hand (YES a raised finger, NO a fist), a live session on the hosted app finished a message in 8 questions and about 2 minutes: "The tube or tape is bothering me.", confirmed before and after it was spoken. It is not a medical device and gives no diagnosis or advice

Team