PicoClaw Claims
Intelligence System

AI-powered triage for Hong Kong Typhoon claims.
Five agents. $100 of hardware. 24/7 coverage.

HK Typhoon Season 2026 Live Prototype Claude AI Powered
The Challenge

HK Typhoon season overwhelms
claims operations

10×
Call volume spike
during Typhoon T10
70%
Customers prefer
phone in emergencies
>60s
Answer speed degrades
from 20s to 60s+
15%
Abandonment rate
surges during T8+

Peak season: May–October. T8+ signals trigger claims surges lasting 24–72 hours.
Human teams simply cannot scale fast enough.

The Solution

Five AI agents. Always on.
Instantly scalable.

📨
Inbound
Message
→
🎯
Triage
Agent
→
💙 Grace
Empathy
⚡ Swift
Fast-track
📚 Kara
Knowledge
🔥 Phoenix
Escalation
→
👤
Human
Queue
→
✅
Resolved

Sentiment-based routing in <3 seconds. Each agent has a distinct personality.
Seamless handoff to human queue when needed. GL8 compliant throughout.

Agent Personalities

Five agents. One brain. Zero wait time.

🎯
Triage
Analyses sentiment & routes instantly. Never speaks to customer.
💙
Grace
Warm, patient, bilingual. For elderly, distressed & injured customers.
⚡
Swift
Efficient & decisive. Rapid property & vehicle damage claims.
📚
Kara
Friendly knowledge expert. Policy questions, FAQ, coverage guidance.
🔥
Phoenix
Calm & authoritative. Escalations, legal threats, complex claims.
👤 Human Queue — Operators monitor in real-time and pick up when AI reaches its limits
Intelligent Routing

Right agent. Every time.

💙 Grace
"scared" · "my grandmother" · "don't know what to do"
"hurt" · "terrified" · "elderly" · distress signals
⚡ Swift
"car crushed" · "roof gone" · "need someone NOW"
property/vehicle damage + urgency signals
📚 Kara
"what does my policy cover" · "what is my excess"
"how do I claim" · "how long will it take"
🔥 Phoenix
"this is unacceptable" · "lawyer" · "sue"
"third time calling" · anger + complexity signals
👤 Human Queue
"speak to a person" · "真人" · fire/gas emergency
After 3 unresolved turns · legal + emotional combo
⚡ Hard overrides (pre-LLM):
"999" · "fire" · "gas leak" → CRITICAL + Grace + alert human
"lawyer" / "sue" → Phoenix instantly
Live Demo

See it in action

💬

Customer Chat

Live AI agents responding to typhoon claims

/chat
🎯

Scenario Simulator

6 typhoon scenarios fire in real-time

/simulator
📊

Live Dashboard

Real-time SLA, tNPS, agent utilisation

/metrics
🌐 https://claims-triage-ai-assistants.vercel.app

🎬 Demo flow: /simulator → Fire "Mrs. Chan — Flooded Flat" → Watch /triage update live → Show /metrics

Management Visibility

Everything management needs. Live.

📊 Service Level Metrics

  • ✅ Total conversations & active now
  • ✅ SLA adherence % (target: >95%)
  • ✅ Average handle time
  • ✅ First contact resolution rate
  • ✅ tNPS (post-chat survey score)

👥 Human Queue

  • ✅ Queue depth & wait times
  • ✅ Who answered, how long
  • ✅ Human vs AI resolution ratio

🤖 Agent Performance

  • ✅ Utilisation per agent (Grace, Swift, Kara…)
  • ✅ Routing accuracy
  • ✅ Avg response time per agent
  • ✅ Escalation rate to human

Target SLAs

Bot response: <3 sec
Agent turn: <30 sec
Human queue: <2 min
Containment: >50%
Hardware Vision — Phased Rollout

Prove it with one board.
Scale to five.

PHASE 1 — ARRIVING
💙
Grace
Board 1 · :8001
PHASE 2
⚡
Swift
Board 2 · :8002
PHASE 3
📚
Kara
Board 3 · :8003
PHASE 4
🔥
Phoenix
Board 4 · :8004
PHASE 5
🎯
Triage
Board 5 · :8005

Mixed mode: Grace runs on hardware. Swift, Kara, Phoenix, Triage stay in simulation.
The dashboard looks identical. Validate first — then add boards one by one.

$20
First board cost
Proof of concept
$100
Full 5-board cluster
Goal state
24/7
Always on
~1–2W per board
Sipeed LicheeRV Nano W

Credit card-sized. Enterprise capable.

📦 Hardware Specs

CPU: SOPHGO SG2002, 1GHz RISC-V
RAM: 256MB DDR3
Net: WiFi 6 + Bluetooth 5.2
Power: USB-C, 1–2W
OS: Buildroot Linux (64-bit)
Size: Credit card

⚙️ Software: PicoClaw

Open-source Go binary that wraps the Anthropic Claude API.

Each board runs picoclaw gateway — an HTTP server with agent-specific personality baked into the system prompt.

🗓 Timeline

Today ✅

Board ordered from AliExpress. SD card flashed. Firmware ready.

19–27 March 2026

Board arrives. Run setup script. Deploy PicoClaw. First agent live.

April 2026

5 boards deployed. Full hardware agent cluster. Pre-typhoon season.

🔌 Migration

Single env var switches between modes:
AGENT_MODE=simulation|hardware

Dashboard is identical. Only the agent calls change — from local functions to HTTP calls to board IPs.

Scalability Story
"During T8, we spun up 10 Grace boards to handle the empathy surge.
Total cost: $200 of hardware.
Human equivalent: $1,000/day."
$0.10
Cost per AI
interaction
$10
Cost per human
interaction
100×
AI vs Human
cost ratio
∞
Scale with
more boards
Cost Analysis

The numbers are compelling

Item Cost Notes
5 × LicheeRV Nano boards $100 One-time hardware cost
Claude API (claude-sonnet) ~$5/day During active typhoon surge
Power (5 boards × 2W) <$1/mo Running 24/7
AI total (surge day) ~$5 Handles 1,000+ interactions
Human agent (surge day) $400–800 8 agents × $50–100/day
AI containment target >50% Reduces human load by half
Security & Compliance

Built secure from day one

✅

GL8 AI Disclosure

Every agent identifies as AI on first contact. Human option always visible.

✅

No Secrets in Code

API keys via environment variables only. .gitignore enforced. Never in repo.

✅

HTTP Security Headers

X-Frame-Options, X-XSS-Protection, Content-Type-Options, Referrer-Policy.

✅

SAST Scanning

ESLint security plugin running on every commit. TypeScript strict mode.

✅

No SQL Injection Surface

In-memory store (no DB). No raw query construction anywhere.

✅

Input Validation

All API routes validate inputs. Hard overrides prevent prompt injection escalation.

⚠️

Auth (Production TODO)

Operator endpoints need auth before production. Prototype deliberately open.

⚠️

Rate Limiting (Production TODO)

Add API rate limiting before production to prevent abuse.

Roadmap

From prototype to production

NOW — PROTOTYPE ✅
  • ✅ 5-agent system running
  • ✅ Sentiment-based routing
  • ✅ Human queue
  • ✅ Live dashboards
  • ✅ Bilingual (EN/ZH)
  • ✅ GL8 compliance
  • ✅ 6 demo scenarios
  • ✅ Security scan
MARCH–APRIL — HARDWARE ROLLOUT
  • 📦 Board 1 arrives (19–27 Mar)
  • 💙 Grace deployed on hardware
  • ✅ Validate mixed mode works
  • 📦 Order boards 2–5 if proven
  • ⚡→📚→🔥→🎯 Roll out agents
  • 📊 Board health dashboard live
  • 🔑 API auth + rate limiting
MAY — TYPHOON SEASON
  • 🌀 Live HK typhoon handling
  • 📞 NICE CXone integration
  • 🗄 Claims Connect linkage
  • 📱 SMS/WhatsApp gateway
  • 🤝 Amelia chatbot handoff
  • 📈 Surge auto-scaling
  • 🎓 Agent self-improvement
Summary

Five agents.
One hundred dollars.
Zero wait time.

Today: running on a laptop, demonstrating the full system.
March: running on $100 of hardware, one board per agent.
May: live for HK typhoon season, before peak hits.

$100
Hardware
<3s
Response time
5
Agents
24/7
Coverage
1 / 14
← → arrow keys to navigate