A browser extension that uses AI to warn people before they send money to a likely scam — reading the live transfer on their bank's page and flagging it the instant something looks wrong.
Built for NexHack 2026 — Track 2: Fintech Risk & Fraud Intelligence.
https://canva.link/cd78rs35llu63hu
- What it is
- Why it matters
- How it works
- Install it (no setup)
- Prerequisites
- Run it locally (for developers)
- Usage
- Configuration
- Project structure
- Testing
- Honest limitations
- Contributing
- Support
- License
Sentinel Scam Shield is a Chrome (Manifest V3) extension plus a small Next.js screening service. When a user is about to send money on a supported bank or e-wallet website, the extension:
- Intercepts the "Send" click before it submits,
- Sends the complete transfer context (recipient, amount, reference/memo) to a multi-agent risk engine,
- Shows a clear warning overlay if the transfer looks like a scam — so the user can cancel before the money leaves.
The engine is not a keyword filter. Three specialist agents screen in parallel — risk (history-free rules), behaviour (recipient history + shared blocklist), and anomaly (population outliers + velocity) — and fuse their signals into a deterministic score. An LLM adjudicator then rules on the evidence and can tip borderline cases, but it is bounded so it can never override strong deterministic signals or blocklist hits.
Authorized-push-payment (APP) scams — fake investments, impersonation, "urgent" transfers — cost Malaysians hundreds of millions a year. The money is gone the moment the victim hits send. Sentinel inserts a last-second, context-aware warning at exactly that moment, on the bank pages people already use, with no integration required from the bank.
┌─────────────────────────────────────────────────────────────┐
│ Bank / wallet website (CIMB Clicks / Maybank2u / etc.) │
│ user clicks “Send money” │
└───────────────────┬─────────────────────────────────────────┘
│ content.js intercepts the click (capture phase)
│ reads payee / amount / memo via a per-bank adapter
▼
┌───────────────┐ ┌──────────────────────────────┐
│ background.js │ ─────▶ │ Next.js POST /api/screen │
└───────────────┘ │ ┌─ main_agent fans out to: │
▲ │ │ • risk_agent │
│ │ │ • behaviour_agent │
│ │ │ • anomaly_agent │
│ │ └─ deterministic fusion ───▶ │
│ │ ┌─ AI adjudicator (bounded) │ → DeepSeek
│ │ └─ state machine │
│ └──────────────┬───────────────┘
│ verdict: allow | warn | block │
◀────────────────────────────────┘
│
allow ───────▶ let the transfer proceed untouched
warn / block ▶ show warning overlay → user cancels or proceeds
The risk engine receives the whole transfer plus any enriched history, so adding a new observed field automatically improves every layer's reasoning. Per-bank field selectors are the only thing that changes between banks, and they live in one file (extension/site_adapters.js).
The screening service is already hosted at next-hack2026.vercel.app, so end users do not need Node, an API key, or a local server. They only load the extension into Chrome:
- Download the extension folder. Either clone the repo, or download it as a ZIP from GitHub (Code → Download ZIP) and unzip it. You only need the
extension/folder. - Open
chrome://extensionsin Chrome (or Edge / Brave). - Toggle Developer mode on — top-right corner.
- Click Load unpacked.
- Select the
extension/folder. - Done — the 🛡️ Sentinel icon appears in your toolbar.
Sharing it: zip the
extension/folder and send it. Whoever unzips it repeats steps 2–5 above. Because the backend is hosted, the zip works for them with zero configuration.
You only need these to run or modify the backend — not to use the extension (see above).
- Node.js 18+ and npm
- A Chromium browser (Chrome, Edge, or Brave) for the extension
- An OpenAI-compatible AI API key — the demo is configured for DeepSeek, but Groq, Moonshot, or OpenAI work by changing three env values. Without a key the app still runs on the deterministic fallback.
# 1. Clone
git clone https://github.com/wenshen18643/nextHack2026.git
cd nextHack2026
# 2. Install dependencies
npm install
# 3. Configure the AI provider
cp .env.example .env
# then open .env and set KIMI_API_KEY to your DeepSeek (or other) key
# 4. Start the screening server
npm run dev # serves http://localhost:3000Load the extension as in Install it. To point it at your local server instead of the hosted one, change default_api_base in extension/background.js back to http://localhost:3000, or set api_base in chrome.storage.sync.
The extension is just a client of one endpoint. You can hit it yourself:
curl -s https://next-hack2026.vercel.app/api/screen \
-H "content-type: application/json" \
-d '{"payee":"MULE HOLDINGS 8829","amount":9000,"memo":""}'Example response (deterministic agents fire even with a blank memo):
{
"advice": "block",
"score": 82,
"state": "DENY",
"reason": "Recipient is on the shared scam-account blocklist, and the amount is far outside normal patterns.",
"signals": [
{ "layer": "behavioral", "code": "KNOWN_FLAGGED_ACCOUNT", "weight": 80, "detail": "Recipient is on the shared scam-account blocklist..." },
{ "layer": "rules", "code": "HIGH_ABSOLUTE_AMOUNT", "weight": 22, "detail": "Large transfer of 9000." }
],
"agents": [
{ "agent": "risk", "signals": [...] },
{ "agent": "behaviour", "signals": [...] },
{ "agent": "anomaly", "signals": [...] }
],
"ai_used": true
}| Field | Meaning |
|---|---|
advice |
allow (let it through), warn, or block |
score |
Risk score 0–100 |
state |
PASS / INSPECT / QUARANTINE / DENY |
reason |
One-sentence explanation shown to the user |
signals |
Every signal that contributed to the score |
agents |
Per-agent breakdown for explainability |
ai_used |
true if the AI adjudicator contributed |
All configuration is environment variables in .env (see .env.example). The AI variables are named KIMI_* for historical reasons but point at any OpenAI-compatible provider:
| Variable | Description | Example |
|---|---|---|
KIMI_API_KEY |
API key for the AI provider | sk-... |
KIMI_BASE_URL |
Provider base URL | https://api.deepseek.com |
KIMI_MODEL |
Model name | deepseek-chat |
KIMI_TEMPERATURE |
Sampling temperature (keep low for reliable JSON) | 0.2 |
To support a new bank, add an adapter to extension/site_adapters.js (its CSS selectors for the recipient field, amount field, and send button) and add the bank's domain to extension/manifest.json.
extension/ # Chrome MV3 extension (load this unpacked)
manifest.json # permissions + which sites to run on
content.js # intercepts the send click, spinner + warning overlay
background.js # calls the screening API
site_adapters.js # per-bank field selectors (the only per-bank file)
overlay.css # warning + spinner styles
popup.html / popup.js # account, settings, and Pro Stripe checkout
src/
app/
api/screen/route.ts # POST /api/screen — the screening endpoint (CORS-enabled)
api/billing/* # Stripe Checkout + webhook + status
api/auth/* # signup / login / logout
page.tsx # landing page
business-case/page.tsx # market evidence, pricing, sizing
lib/agents/
main_agent.ts # orchestrates the specialist agents + AI adjudicator
risk_agent.ts # history-free rules (amount, timing, text)
behaviour_agent.ts # recipient history + shared scam blocklist
anomaly_agent.ts # population outliers + velocity
types.ts # TransferContext, AgentReport
lib/screen/
ai_screener.ts # LLM adjudicator (full-context prompt)
service.ts # API entry point that builds context and runs main_agent
cold_rules.ts # deterministic history-free rules
lib/risk/ # score fusion, state machine, shared types
lib/billing/ # Stripe client and helpers
lib/db/ # Supabase client + RPCs
lib/auth/ # session + profile helpers
docs/ # judge-ready product, design, engineering, business docs
npm test # Vitest unit tests
npm run typecheck # TypeScript, no emit
npm run lint # ESLint (next lint)
npm run build # production buildThese are deliberate, and worth stating plainly:
- It warns, it does not freeze. No third party can stop another bank's transfer; doing so needs a banking license. Sentinel only advises — which is what makes it adoptable without regulatory approval.
- It works on bank websites, not app-only wallets. A browser extension can reach CIMB Clicks on the web, but not the Touch 'n Go app. App coverage would need a separate Android accessibility build.
- No transaction history. On a real bank page the extension sees only the current transfer, so it does scam-pattern detection on that transfer rather than behavioral profiling — which is exactly what catches APP scams.
- Demo-grade extension. Bank pages change their markup and have anti-tampering defenses; the production model is an API a bank embeds, not a sideloaded extension.
Contributions are welcome:
- Fork the repo and create a feature branch (
git checkout -b feat/my-change). - Make your change; ensure
npm run typecheck,npm run lint, andnpm testall pass. - Open a pull request describing the change and how you verified it.
For a new bank adapter, include the field selectors and a note on how you confirmed them.
- Bugs / feature requests: open a GitHub issue.
Released under the MIT License.