Auntie AI · Before You Click

A senior-friendly AI co-pilot that catches scams before money is lost. Built for the Global AI Hackathon Singapore 2026 (CodeBuddy × GLM).

Build started: 2026-04-25 · Solo MVP · 5-day sprint

🌐 Live Demo

👉 Try Auntie AI now — deployed on Vercel 👉 Try Auntie AI now — deployed on CloudStudio

---

What This Is

When a senior receives a suspicious SMS, payment request, or investment invite, they tap one button. Auntie AI:

1. THINK — reads the screen content with GLM-4.5V (multimodal vision) or GLM-4.5 (text)
2. DECIDE — classifies risk across 7 scam dimensions, assigns Safe / Low / Medium / High / Very High
3. EXECUTE — generates a plain-language explanation, recommended safe actions, and (for High+) a redacted message draft to share with family via WhatsApp

The whole agent loop runs in 5–15 seconds and is visible to judges on Demo Day.

---

Quick Start

# 1. Install dependencies
npm install

# 2. Set up environment
cp .env.example .env.local
# Then edit .env.local and paste your ZAI_API_KEY

# 3. Run dev server
npm run dev
# Open http://localhost:3000

Click any of the 5 sample scenarios on the home page — they call the real Z.ai GLM API and return a structured risk assessment.

---

What Day 1 Ships

File	Status	Purpose
app/api/check/route.ts	✅	Real Z.ai GLM-4.5 call returning validated JSON
lib/prompts.ts	✅	Tuned system prompt with 3 few-shot examples
lib/scenarios.ts	✅	5 demo scenarios (bill + 4 scam types)
lib/types.ts	✅	Zod schema for RiskAssessment
lib/i18n.ts	✅	en/zh translations
lib/store.ts	✅	Zustand store + useT() hook
app/page.tsx	✅	Home page with end-to-end demo flow
components/RiskCard.tsx	✅	Result card with risk badge, signs, actions
components/AgentSteps.tsx	✅	Think · Decide · Execute loading state
components/LangToggle.tsx	✅	EN/中 switch

---

Day 2 Features (All Completed ✅)

Feature	Prompt	Status
Hokkien warning audio (play/pause/auto-play)	#1	✅
Upload page (drag-drop + camera + sample grid)	#2	✅
Thinking page (agent steps + API call)	#3	✅
Result page (RiskCard + FamilyAlertCallout)	#4	✅
Family Alert page (recipient card + copy + WhatsApp)	#5	✅
Vision API (multimodal screenshot analysis)	#6	✅
Ambiguous case handling (4 few-shot examples)	#7	✅
Live deployment + smoke test	#8	✅

---

Architecture

Home (/) → Upload (/upload) or pick a scenario
        ↓
Upload: drag-drop/camera → FileReader → base64 dataURL in Zustand
        ↓
Thinking (/thinking) — shows AgentSteps (THINK · DECIDE · EXECUTE)
        ↓
POST /api/check { screenText OR imageDataUrl, language }
        ↓
route.ts routes to assessText() or assessImage() (multimodal vision)
        ↓
OpenAI gpt-4o-mini with tuned system prompt + 4 few-shot examples
        ↓
Zod validates RiskAssessment JSON → store.setAssessment()
        ↓
Result (/result) — <RiskCard> + <FamilyAlertCallout> + audio button
        ↓
[High/Very High] Alert (/alert) — recipient card, copy, WhatsApp CTA

---

Testing the GLM Call Manually

curl -X POST http://localhost:3000/api/check \
  -H "Content-Type: application/json" \
  -d '{
    "screenText": "Your DBS account will be suspended. Click bit.ly/x to verify in 10 min.",
    "language": "en"
  }'

Expect a response like:

{
  "ok": true,
  "assessment": {
    "riskLevel": "Very High",
    "scamType": "bank-phishing",
    "summary": "This is a scam. Do not click the link.",
    "warningSigns": ["10-minute deadline", "shortened bit.ly link", "account threat"],
    "plainExplanation": "Real banks do not ask you to verify through random short links...",
    "recommendedActions": ["Do not click the link", "..."],
    "familyAlertDraft": "Hi 小芳, mum just got an SMS...",
    "confidence": "High",
    "language": "en"
  }
}

---

Submission Checklist

• CodeBuddy deployment URL — Live Demo
• Vercel backup URL (insurance)
• 90-second demo video
• 7-slide pitch deck
• GitHub repo public + clean commit history starting 2026-04-25
• README first line shows date stamp

---

Tech Stack

• Next.js 14 (App Router) + TypeScript + Tailwind CSS
• Zustand for client state
• Zod for runtime schema validation
• OpenAI gpt-4o-mini (text + vision/multimodal) — API key required
• CloudStudio for deployment

---

License

Copyright (c) 2026 Elena. All rights reserved.

This source code is shared publicly for educational and competitive review purposes only (CodeBuddy × GLM Global AI Hackathon Singapore 2026). Commercial use, redistribution, or derivative products require written permission. Contact: leini8891@gmail.com