The problem SHOU_Team_Web_1.5 solves

She isn't hacked. She's talked into sending it herself — "manufactured consent."

RM2.97 billion lost to scams in just 11 months of 2025 (67,735 cases). Bank Negara Malaysia's 2025 Annual Report puts 95% of Malaysia's online fraud cases as "authorised transactions" — the victim approved the transfer herself, no key stolen. Bukit Aman CCID logged RM3.11 billion in total commercial crime losses for 2024 alone, RM2.45 billion of it fraud across 37,240 investigation papers. Seniors are the sharpest edge of that curve: only 6.4% of victims (5,533 of 86,266, 2021–2023) but 20% of total losses (RM552.5M of RM2.7B) — and by mid-2024 that gap had widened to 8.3% of victims vs. 27.7% of losses.

RM1.47 billion from investment scams alone — over half of all losses. Investment scam losses jumped from RM848.6M (6,337 cases) in 2024 to RM1.47B (9,603 cases) in 2025, a ~73% year-over-year rise, with syndicates moving victims from Facebook/ Instagram into WhatsApp or Telegram "trading groups" before blocking withdrawals. Love scams cost RM45.9M in 2024 (up from RM43.9M in 2023), 75.6% of victims women 61+, and 370 cases started on Facebook, 332 on WhatsApp in 2023 — the two most common contact points. The US shows the same shape: FBI IC3 logged 147,000+ complaints from adults 60+ in 2024 — more than any other age group — for $4.8–4.9B, up 43% YoY, averaging $83,000 per victim.

70% of victims never report — real losses may hit RM54 billion. And what does get spent goes to the wrong category: banks blocked RM383M in fraudulent transactions in 2023 alone (RM780M+ combined 2023–2024), and the National Scam Response Centre froze 162,642 mule accounts in the latest reporting year — none of it touching the 95% "authorised transaction" category, because it's all built to catch stolen credentials, not manufactured consent. Once sent, instant payment rails move funds irreversibly within seconds, and no existing tool sits at that exact point: ScamShield/Truecaller warn on known numbers but go silent once a chat moves to WhatsApp; BioCatch reads banking-session hesitation, not the conversation that caused it. Nobody else sits in the top-right — blocking it before, at the moment money is given away.

Challenges we ran into

1. What if the AI gets tricked? (Math over promises)
   The Challenge: If a scammer jailbreaks the LLM or it hallucinates and calls a real scam "LOW risk", a wallet that blindly trusts AI loses her life savings.
   How We Overcame It: We enforced safety in Sui Move with one line of math: effective_tier = max(amount_tier, claimed_tier). Even if the AI claims a $600 transfer is completely safe, the contract independently detects it crossed her personal ceiling and forces it to HIGH escrow anyway. The chain never has to trust the AI to protect her.

2. Real-world AI infrastructure is unpredictable
   The Challenge: Decentralized AI inference has real-world latency spikes (measuring up to 26s+ on cold calls and 524 timeouts on Kimi model). We couldn't let network lag leave her unprotected during a live scam.
   How We Overcame It: We decoupled safety from AI availability. Instant deterministic threat floors (authority impersonation + urgency + secrecy demands) run locally on-device. If Gonka models lag, disagree, or go offline, the hard floor automatically takes over and holds the funds. Safety never depends on Wi-Fi holding up.

3. Passive chat scanning vs. Spyware & Family abuse
   The Challenge: Passively monitoring WhatsApp sounds like spyware. Worse, family members are statistically a leading vector for elder financial abuse—handing her son her chat transcripts would just replace an online scammer with family coercion.
   How We Overcame It: We built structural privacy. Identifying PII (names, IC, phone numbers) is stripped on-device before transmission. Scoring runs inside a confidential enclave that only emits an opaque hash and signature. And the Guardian Dashboard only shows transfer amounts and recipients—never her private conversations.

. "Just push to Vercel" broke our security architecture
The Challenge: Serverless platforms wipe in-memory state on cold starts. Our enclave holds an active signing key in memory; a cold start would mint a new key mid-demo and instantly invalidate every on-chain attestation.
How We Overcame It: We rejected the serverless shortcut. We containerized our persistent Node services with Docker and deployed on Railway, preserving in-memory enclave keys, loopback network security, and cryptographic integrity with zero broken sessions.