LifeLink

One calm, voice-first place for older adults to stay connected, get through the day, and stop scams, with someone they trust one tap away.

Speak or type → review the exact action → approve → done.

Try it live at lifel.ink · Devpost

Built in 36 hours at SBUHacks 2026 (Stony Brook University, October 23 to 25), theme Steel and Steam.

---

Contents

The project

• Why we built it
• What it does
• The AI never acts on its own
• How we built it
• Try it
• Challenges, wins, and what's next

For developers

• Getting started
• Environment
• Product
• Design
• Architecture
• Integrations
• Development
• Project status

---

Why we built it

Scammers target older adults online all the time. They buy ads and send emails pretending to be Medicare or a bank. Last month it happened to a teammate's grandmother: she spent four hours on the phone with a scammer and lost over $90,000.

The usual advice is to let family handle everything, and she refused. She wanted to keep doing things herself. So we built one interface where older adults act for themselves in a controlled, safe environment, and where someone they trust is a tap away, so nobody spends hours talking to the wrong person.

We aren't the only ones who see the gap. Y Combinator's Fall 2026 Request for Startups includes AI for the Aging Population. It notes that by 2030 one in five Americans will be over 65, that about 53 million family members already provide unpaid care, and that few technologies, even popular voice assistants, are designed for older adults. It calls out the same areas LifeLink focuses on: conversational voice interfaces, safety and independence, and software that helps families coordinate care and appointments.

What it does

Stop scams

• Every new email is screened and labelled Likely a scam, Warning signs, or No warning signs, with the exact reasons, so people learn the patterns as well as get the verdict.
• Check anything: paste a text, photograph a letter, pick a screenshot, or put a call on speakerphone for live captions and a spoken warning when a scam is detected.
• Ask someone you trust: when a scam is detected, one tap prepares a message to someone in your Trusted Circle. It still needs your approval to send. Family you invite can see scam alerts without ever seeing the message itself.
• Help me understand explains any confusing message or email: what it's asking for, the dates, what's missing, and a few replies you could send.

Stay connected

• Message family over iMessage, email, Discord, or Telegram from one simple interface, protected by the same scam checks. Send one message to several people across platforms at once ("Tell all my grandkids I'd love to see them this weekend").
• Ask "Can Sam drive me to my appointment Tuesday?" LifeLink writes the message. When Sam answers, it tells you whether that's a yes, a no, or a different time, then offers the calendar event and a reminder.
• See photos and documents that friends and family send.
• Keep an eye on stocks and commodities for retirement savings.

Your day

• Read my day reads the weather aloud, says who wrote to you, and flags any email that looks like a scam, so nothing comes as a surprise.
• Weather next to each appointment shows when rain or ice might change plans.
• News without the ads: uplifting stories from Good News Network and Smithsonian, or world news from NPR. Tap a headline for a short summary inside the app. No outside sites and no endless feed.
• How I like help: short or detailed answers, one step at a time, ten languages (English, Spanish, French, German, Italian, Portuguese, Simplified Chinese, Vietnamese, Korean, Tagalog), theme, font, text size, and read-aloud voice.

Groceries

• Paste a list from a text or email and every item is added at once.
• Items are sorted by aisle in the order you walk the store, with cold and frozen last.
• Set a budget, see the estimated total, save a usual list, reuse past trips, and print or copy the list for whoever's shopping.

Fun

• Solitaire and Mahjong with no ads and no sign-up. Both work offline.

Everything works with read-aloud, large text, high contrast, full keyboard control, and all ten languages. Practice mode lets anyone try every feature on sample data first, with nothing real sent.

The AI never acts on its own

The model only proposes. The Convex backend writes out the exact action, fingerprints it, and runs it only when the approval matches the current version and the permissions still hold:

$$ \text{execute} \iff H(d_{\text{approved}}) = H(d_{\text{current}}) \ \land\ \text{permissions still hold} $$

If anything changes after you approve (the wording, the recipient, the day), the old approval no longer counts and you approve again. Family members only ever see what you chose to share. They can never act for you, and you can revoke their access instantly. The architecture section lists the guarantees the backend enforces and tests.

How we built it

Stack: Next.js 16 (React 19, Tailwind v4, installable PWA), Convex, and Better Auth in a Bun + Turborepo monorepo. The web app is deployed on Vercel. The Discord bridge holds a persistent Gateway connection, which serverless hosting can't keep open, so it runs as a systemd service on a Vultr Debian VPS.

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗