# i-persons

- **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:** [Mohamed Amri](https://cerebralvalley.ai/u/WorldClock)
- **GitHub:** https://github.com/mohamedamrimail/metamusehackathon
- **Demo video:** https://www.youtube.com/watch?v=jNVmFmvy1KM
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1938

i-persons: your living digital twin, built on Meta's Muse models.
Your i-person knows you, speaks for you, and acts for you, but never without your yes.

Live: https://i-persons.com (beta)
Code: https://github.com/mohamedamrimail/metamusehackathon
Demo video: https://www.youtube.com/watch?v=jNVmFmvy1KM
By: Mohamed Amri, Yousoft-IT (registered company in Sfax, Tunisia, with a verified Meta Business portfolio; https://yousoft-it.com)
All code in this project was written during the hackathon.

THE PROBLEM
People lose hours every week to small coordination: finding a time for a coffee with family, booking a meeting, reordering the usual things, answering the same questions about themselves. Assistants today either just chat, or act without asking. People want the help, but not at the cost of control or their private data.

THE SOLUTION
i-persons gives every person a digital twin that takes this work over while the person keeps full control. It acts for you, learns you and speaks for you, with one hard rule: nothing happens in your name without your yes.

WHO IT IS FOR FIRST
Families who coordinate across generations and languages: a daughter in one city, her mother in another, a grandmother who only speaks Tunisian Arabic. Their twins plan together and every person approves with one tap, on the web or in WhatsApp, which they already use.

WHAT IT DOES
1. A living twin from one selfie. You take a selfie and confirm "This is me". Muse Image turns it into your twin in the style you pick and makes its poses (smile, blink, wave, thinking). SAM 3.1 finds the person, face, eyes and mouth in each pose, so the twin breathes, blinks, lip-syncs and changes mood with every reply. People who prefer not to use a selfie pick one of 72 neutral avatars.

2. It acts for you, always with your yes. Hold the mic and talk: Muse Voice Transcribe turns speech into text in about 0.2 s, and the twin answers out loud in English, French, German, Arabic or Tunisian Derja. Muse Spark understands the request, uses tools (calendar, contacts, web search, orders and bookings) and prepares the action. Every action that changes something becomes an Approve / Decline card, on the web and as WhatsApp buttons. Your own rules ("never book meetings before 9 am") are checked by Muse Spark on every action; a proposal that breaks a rule never appears. For hotels, flights and shopping, the twin prepares real options and a prefilled link; it never buys anything by itself.

3. Twin to twin. Sara says "Plan a coffee with my mom next week." Her twin asks her mother Hana's twin. The two twins compare only free times, never what is in the calendars, and agree on three slots and a café. Hana gets a card; after both say yes, the coffee is in both calendars. Twins connect only when both people agree, and either can disconnect at any time.

4. It learns you, under your control. The twin remembers routines, goals and preferences; every fact is visible, editable and private by default. A read-only Outlook or Gmail connection, switched on per service, turns booking and order emails into insights such as "you book hotels about 7 days ahead"; the emails themselves are never kept. "Time given back" shows the time saved, counted only from actions you approved.

5. It speaks for you. At i-persons.com/@you, visitors can talk to your public twin without an account. It says it isn't you and only uses the facts you made public; messages and meeting requests become cards for you. Companies connect through a Partner API (pseudonymous ids, matching on public facts only, never gender or age) and a no-code partner console.

6. Everywhere you are: the web, a home-screen app, WhatsApp (text, voice notes, pictures, approve buttons) and Meta Ray-Ban Display glasses, where the twin floats in view and you approve cards with the Neural Band.

7. A door to many services. Through the Partner API, your twin talks to services for you, shares only what you allow, and brings their answers back as cards for your yes. Today: shopping, hotels and flights (real options with prefilled links) and job search (recruiters reach your public twin through the partner console). Next: education, health and more service partners. The twin carries your request and your preferences; each service brings its own expertise, and nothing is shared, booked or bought without your yes.

SAFE BY DESIGN
- The approval gate is enforced in the runtime, not in the prompt: tools that write or spend money can only propose a card.
- Rules are checked on every action, and the check fails closed.
- Everything from outside (web, emails, visitors, partners) reaches the model as untrusted data, and each card says what it was based on.
- A fact learned right after reading outside content waits for your "Keep".

WHAT WORKS TODAY
Live in production at https://i-persons.com (beta, new accounts approved by the team), and runnable locally with "docker compose up --build" and only a Meta API key. Working end to end: the setup flow, the living avatar, push-to-talk voice in five languages, approval cards on the web and WhatsApp, rules on every action, prompt-injection protection, twin-to-twin planning into both calendars, mailbox insights, public twins, the Partner API, WhatsApp and Meta Ray-Ban Display.

READY FOR PRODUCTION
- Business side already running: AI credit metered at Meta's price, gifts and a family plan, beta approval, a country allowlist with a waitlist, and an admin panel with metrics and feedback.
- Measured against the real Muse Spark (muse-spark-1.3), not mocks: a 50-case agent eval in five languages scored 100% (50/50); rules kept 5/5; prompt injections acted on 0/5; private-data leaks to visitors 0/5. The first run scored 96% and found a real bug, now fixed.
- Voice latency (p50, 30 real spoken turns): 0.18 s from end of speech to transcript, 2.29 s to the reply, including tools and the rules check.
- About 21,700 lines of strict TypeScript and 319 automated tests.
- Median AI cost per active user per day: about $0.07.
- Traction is still early: the app opened to the public days before the deadline, with 2 sign-ups and 1 active user as of Nov 1.

TRY IT
Your own account, no beta wait: https://i-persons.com/?invite=TRY-IPERSONS
Demo accounts (password: ipersons-demo):
- sara@i-persons.com: hold the mic and say "Plan a coffee with my mom next week"
- hana@i-persons.com: approve Sara's coffee plan; see insights, rules and time given back
- adam@i-persons.com: My twin > Your look > New selfie; on Talk, "Book Thursday at ten with the supplier"
- giga@i-persons.com: talk in Tunisian Arabic
Visitor view of a twin: https://i-persons.com/@hana
Demo accounts show a live "Meta AI calls" panel listing every Meta model call with what it did and its latency.

WHAT'S NEXT
Your own voice for your twin (cloned with recorded consent), a native glasses app with Meta's Wearables Device Access Toolkit so the twin hears through the glasses and sees what you see, payments for credit top-ups and partner billing, an MCP server for partners' own agents, and a larger eval in more languages, rerun on every release.

## More from Meta Global AI Developer Hackathon

- [Canyouseeme](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1935)
- [Mothertongue](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1936)
- [Quill: recovering handwritten weather records](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1937)
- [WordBlade](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1939)
- [Senti_One](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1940)
- [Team Ayush](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1941)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1938. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt