CrowdShield

The bank sees the transfer. The crowd saw the call. CrowdShield joins them in one scam graph, with NVIDIA Nemotron, to stop scam payments in Malaysia before the money moves.

Comment

CrowdShield

The bank sees the transfer. The crowd saw the call. CrowdShield stops the scam payment before the money moves.

Try it: https://crowdshield.184-197-8-50.sslip.io · Telegram @CrowdShieldBot · banking-app demo at /demo/bank-app · portal logins [email protected] / [email protected], password crowdshield-demo (synthetic data only)

Inspiration

From January to November 2025, Malaysian police recorded 67,735 online crime cases with losses over RM2.7 billion. Scam calls were the biggest category. The national scam hotline, NSRC 997, saved about RM34.5 million in 2025. Most money is never recovered once it has left the victim's account.

Every scam has a moment when it could be stopped, but the signal is split. The victim knows a "police officer" is on the phone. The bank knows a large DuitNow transfer is going to an account opened nine days ago. The business knows its supplier never changed banks. Nobody sees all of it. We built the shared picture.

What it does

CrowdShield is one scam graph with five front doors:

  1. Telegram bot (the public). Forward a number, account, link, message or screenshot. You get a verdict (confirmed scam, high risk, reported, no reports found) with its reasons and one plain action, in Malay or English. Nemotron reads screenshots and names the scam pattern; known scam waves are flagged. If you lost money, the bot walks you through a report and hands you a pre-filled sheet to read from while calling NSRC 997.
  2. Bank portal (fraud analysts and supervisors). Payee checks that combine the payee's risk with the customer's live call signal, an alert queue with graded interventions, cases with timelines and money trails, an AI case brief with a referral note for NSRC/PDRM, and AI-detected scam campaigns for a person to confirm.
  3. The customer's banking app. Before a DuitNow transfer, the bank's backend calls CrowdShield with its partner key. A transfer to a mule account during a call from a confirmed scam number is paused, and the customer sees one clear warning.
  4. CrowdShield Guard, a Chrome extension. Most victims only realise after the fact, and saving a screenshot and sending it somewhere is friction. With the extension you press Alt+Shift+S and drag over any message, email or chat; Nemotron reads it and the graph checks every number in it. On WhatsApp Web, Telegram Web and Gmail (each turned on by the user), scam-linked numbers and accounts are highlighted as they appear, and "it's me, new number, please transfer" impersonation messages are flagged before anyone pays. Detection runs on the device; only the identifiers are looked up.
  5. Supplier check (small businesses). Upload an invoice. Nemotron reads it; a changed bank account is caught and the payment held; the supplier's official phone number is found on the web for a call-back. Rejecting a fake invoice reports the account, protecting the next business.

Every report, bank confirmation and rejected invoice strengthens the graph for everyone.

How we built it

  • NVIDIA Nemotron on Nebius Token Factory for every model task:
    • Nemotron 3 Ultra classifies forwarded messages in Malay, English, mixed Malay-English and Chinese, judges web search results and link pages, and, with reasoning on, writes case briefs and labels new scam campaigns.
    • Nemotron 3.5 Super VL reads scam screenshots and invoice PDFs verbatim (digits and account numbers exact).
    • Nemotron 3 Super sorts this week's scam news.
    • Qwen3-Embedding-8B embeds scam scripts into pgvector for campaign matching and detection.
  • Model choices are measured, not guessed. We built a 41-message gold set (four languages, 12 scam types plus normal messages) and evaluated Lightning, Nano, Super and Ultra, with reasoning on and off. Defining every scam type in the prompt took accuracy from 68% to 100%; strict JSON-schema decoding stayed on (relaxing it broke 40 of 41 outputs); reasoning didn't help classification, so it is only on where the task needs it. Ultra was the most accurate and the fastest, so it handles interactive text.
  • Tavily for live web intelligence: what the scam graph hasn't seen yet (see "How we use Tavily" below).
  • Deterministic core. Scores come from the graph: noisy-OR over independent sources, half-lives, Sybil-resistant reporter weighting, one-hop propagation along co-case and money-flow edges. It is a TypeScript port of our Python reference, checked by a 500-case parity test. No model sits in the payment path; a payee check takes about 5 ms.
  • Every AI result shows its trace: model, reasoning on or off, prompt version, time and tokens.
  • Stack: NestJS API with one typed contract (Zod) shared by the API, web app and bot; Next.js 15; grammY; FastAPI gateway; Postgres + pgvector; Redis; S3; Docker on AWS with Caddy HTTPS. 140 TypeScript and 12 Python tests run in CI on every push; 11 browser tests (including the extension in a real Chromium) run against the live stack.

How we use Tavily

A brand-new scam number, link or company has no reports in our graph yet. Tavily is how CrowdShield still says something useful on day one. Tavily finds and fetches; Nemotron reads and judges; the graph decides.

When Tavily Nemotron What the user gets
A number, account or link has no reports in the graph Search for it in its local formats ("011-23454821" OR "01123454821" + scam words) Keeps only results that really mention it as a scam A follow-up with cited links, marked "leads, not confirmation"
Someone forwards a link Extract: Tavily opens the page, so the user's phone never does Does it impersonate a brand? Ask for a login, OTP or card? Offer an APK? "Looks like a fake LHDN page. Don't open it."
/invest <platform> Search restricted to the regulators' sites (sc.com.my, bnm.gov.my) plus a news search Keeps results about that platform Regulator alert-list and news hits, never "this is safe"
A business checks a supplier Search and extract the company's contact page Takes the phone number only if the page names the company "Call the supplier back on 03-…, from their own website"
The landing page's scam news News search, last 7 days, in Malay and English Keeps real Malaysian cases, removes duplicates, extracts loss and state "Scams in the Malaysian news", every item linked to its source

Guard rails:

  • Web results are leads with citations, never evidence: they don't change a risk score.
  • A lead counts only if the page actually mentions the identifier.
  • Official bank and government domains are never flagged as impersonations.
  • Only the identifier or a public company name is sent to Tavily, never the user's message.
  • Results are cached (24 h, news 6 h).
  • If Tavily is slow or down, the graph's answer still arrives instantly; only the follow-up is missing.

Example in Telegram

The bot answers from the graph in under a second, then sends what the web says a few seconds later. Synthetic identifiers; the web results here are illustrative.

You:   Someone WhatsApped me from 013-888 2741 offering a part-time job.
       They sent this link: lhdn-refund-my.com/claim

Bot:   🔍 NO REPORTS FOUND
       +60 13-***-2741
       What to do: No reports found, but that doesn't mean it is safe.
       Police and BNM never ask you to transfer money.
       [🚩 Report this] [📞 Call 997] [❓ Why this result?]

Bot:   ⚠️ HIGH RISK
       lhdn-refund-my.com/…
       14 report(s) from the public
       Campaign: Fake LHDN tax refund call + link, Oct 2026
       What to do: Don't open the link or enter your banking login or OTP.

       … a few seconds later (Tavily + Nemotron) …

Bot:   🌐 Web check +60 13-***-2741: found 2 mention(s) online.
       These are leads, not confirmation:
       • "Beware 013-888 2741, part-time job, asked for a deposit" (forum)
       • "List of part-time job scam numbers" (community list)

Bot:   🔗 Link check (we opened it safely, so you don't have to)
       lhdn-refund-my.com/…
       ⚠️ Looks like a fake LHDN page.
       It asks for a login, OTP or card details.
       Don't open it.

You:   /invest Zentrova Capital

Bot:   📈 Investment check: Zentrova Capital
       nothing found on the SC / BNM alert lists or in the news. That isn't
       approval: check that the company is licensed on sc.com.my before investing.

The number nobody had reported now comes with two cited warnings from the web, the link is checked without the user opening it, and an unknown investment platform is never called safe.

Privacy and fairness

Identifiers are looked up by keyed hash and stored encrypted; everything on screen is masked (+60 11-***-4821). IC, phone and account numbers are redacted before any text model sees them. Reporters are never named. An unknown number gets "no reports found", never a green tick. Listed owners can appeal; a supervisor decides; an upheld appeal clears the score. Official numbers (PDRM, BNM, banks) are protected and shown as "may be spoofed". Every lookup and decision is audited.

Challenges

  • Wrong labels came from the prompt, not the model. Every model made the same mistakes until we found that the prompt defined only 5 of the 12 scam types.
  • Cross-language campaign matching. Malay and English reports of the same scam score around 0.72 similarity; we calibrated thresholds on real embeddings (same scam 0.65–0.80, different scams 0.26–0.66) and require 3+ distinct reports before proposing a campaign.
  • Web search false positives. Names in our demo data first matched real businesses, so every name is coined and checked. A lead counts only if the page actually mentions the identifier, and official domains are never flagged.
  • Reasoning budgets. Reasoning tokens count against the output limit; long case briefs were cut off until we raised the budget and asked for concise points.

Accomplishments

  • A transfer to a mule account during a live scam call is paused in about 5 ms, with no model in the path.
  • A new scam wave (fake "KWSP Account 3 agents") is detected from three reports in two languages and labelled by Nemotron Ultra with a summary, repeated phrases and a draft warning; the bot warns the public once an analyst confirms it.
  • An Ultra case brief turns a timeline and money trail into a cited brief and a bilingual referral note in about 12 seconds.

What we learned

Measure before choosing a model; the cheapest fix is often the prompt. Keep models where they are strong (reading, classifying, drafting) and keep decisions in code that can be explained and audited.

What's next

Integrating with real banks. In production CrowdShield advises and the bank decides: it returns a score, reasons and a recommended action, and the bank's own fraud engine makes the call, as banks already do with other third-party risk signals. We see three ways in, from most to least trust required:

  1. Live API (what the demo shows): the bank's backend calls the payee check before releasing a transfer.
  2. Through PayNet's National Fraud Portal: CrowdShield's signals reach banks through infrastructure they already trust. This is the most realistic path in Malaysia.
  3. Data feed: the bank downloads risky identifiers and scores them inside its own systems, so no customer data leaves the bank.

For a pilot, that means:

  • hosting in Malaysia (the demo runs in AWS us-east-1);
  • per-bank credentials with mutual TLS;
  • ISO 27001 or SOC 2 and independent penetration tests;
  • a strict timeout, so a slow check never blocks a payment;
  • each bank reporting only its own side of a money trail, with alerts going straight to the bank holding a traced account;
  • a legal review of banking secrecy and outsourcing rules with the bank and Bank Negara Malaysia.

On iPhones, a banking app can tell that a call is active but not who is calling, so the caller's identity would come from victims' reports or a telco partnership.

More ways in for the public:

  • the Chrome Web Store listing for the extension;
  • WhatsApp (the bot's conversation is already channel-agnostic);
  • an Android companion that checks incoming SMS and WhatsApp notifications on the device, with the same on-device detection as the extension;
  • voice-note transcription when a speech model is available on Token Factory.

Partners: a pilot with one bank's fraud team using real call-signal consent, and case intake from NSRC via API.

Built With

Share this project:

Updates

Submission history