# MomenTalk

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Ahammad Shawki](https://cerebralvalley.ai/u/ahammadshawki8)
- **GitHub:** https://github.com/ahammadshawki8/Momentalk
- **Demo video:** https://youtu.be/xKECwSbhGJU
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3997

Every moment has a story. Momentalk turns everyday photos into a voice for those who cannot rely on speech, making the world around them easier to express.

Raya is five. She understands almost everything around her, but she cannot rely on speech to say it. Children like her, and adults after a stroke, communicate with AAC: pictures that speak for them. Some of the best evidence for young children supports visual scene displays, real photos of their own life where touching Grandma says "Grandma". A 2024 synthesis found 12 of 14 studies with young autistic children met every quality indicator and showed positive results.

The problem is setup. In today's apps a caregiver traces every outline and types every phrase by hand, so the moments that matter rarely get a voice.

Momentalk removes that work with Meta's models:

- Snap a moment: Muse Spark understands the photo and decides what matters. SAM 3.1 outlines every person and object to the pixel, and each becomes a button that speaks in the child's voice, in about 20 seconds. Tap anything missed and SAM finds it from one point.
- Talking videos: SAM 3.1 tracks people and pets through a clip, so outlines follow them as they move.
- Live Modeling: when a parent talks, Muse Voice Transcribe streams their words and matching outlines light up, so the child sees the word while hearing it. Aided language modeling, done automatically.
- Say who's in it: the parent describes the photo aloud and generic labels become names ("big dog" becomes Biscuit).
- Moment Builder: a Muse Wonder agent turns the day's photos and videos into draft moments and suggests words to model from tap history. Nothing reaches the child until a caregiver approves.
- Wish scenes: Muse Image draws the child's illustrated avatar into places they want to go, so they can ask. The avatar is built from a description, never a photo of the child.

The child's view is full screen: big words, instant speech, and an exit only an adult can hold. The caregiver side is a calm, mobile first app with onboarding, a built in camera for photos and video, a words library, and a review queue.

Under the hood: React and FastAPI, SAM masks decoded with Meta's meta-sam-parser, a realtime speech proxy that keeps keys server side, a tool calling agent with a step limit, a spend cap, and Postgres persistence. Voices use free edge-tts, since Meta has no text to speech model.

It runs in any phone browser, in the family's own language, with no special device. Momentalk is a communication tool, not a medical product. Everything was built during the hackathon. Try it live: https://momentalk.onrender.com

## More from Meta Global AI Developer Hackathon

- [ShopFloor Copilot](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3994)
- [Avinash Sharma](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3995)
- [Hedgeling](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3996)
- [Aera](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3998)
- [BeforeTheTape](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3999)
- [HerPattern Pocket](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4000)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3997. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt