Beacon: Scam Message Checker for Malaysia

Overview

Beacon exists because the people most targeted by scams are the least equipped to spot them. Someone pastes or speaks a message and gets one of three plain answers, Looks okay, Be careful, or Likely a scam, with the reasons in their own language and what to do next. A single Jev call answers seven questions about the message in parallel: an overall verdict plus six warning signs, returned as typed probabilities rather than text to be parsed. The result rules lean careful on purpose, so a strong warning sign or low model confidence can never produce a safe verdict. When Jev cannot be reached, a built-in check still answers and says plainly that it is less accurate. Every scam and caution result carries a report step that asks one question, whether money or credentials are already gone, and routes accordingly: tap-to-call 997 at the National Scam Response Centre, the MCMC complaint form, or PDRM CCID's Semak Mule to check a number or account, with a copyable report summary in the reader's own language. Beacon files nothing itself, because no Malaysian agency accepts a report by API, and nothing leaves the device. The design follows the audience: 50px tap targets, three text sizes, every text and button pair at WCAG AA or better in both themes, speech input and read-aloud, and a page that opens with no connection. On Android it registers as a share target, so a scam message goes from WhatsApp to an answer in two taps.

Key features

• One Jev call answers seven questions about a message in parallel and returns typed, calibrated probabilities rather than prose to parse
• Four languages, English, Bahasa Melayu, Chinese and Tamil, starting in the phone's own language, covering the answer, the reasons and the advice
• Guided reporting: one question routes the reader to 997, the MCMC complaint form, or PDRM CCID's Semak Mule, with a copyable report summary in their language
• Result rules that lean careful: a strong warning sign or low model confidence can never produce a safe verdict
• Works with no connection. A built-in check answers when Jev is unreachable and says plainly that it is less accurate
• Built for the people who need it: 50px tap targets, three text sizes, speech input, read-aloud, WCAG AA contrast in light and dark
• Android share target, so a scam message goes from WhatsApp to an answer in two taps, and installable on every platform
• Privacy by design: the message is never stored, only the phone numbers and links found inside a scam, and only when recording is switched on
• Verified by more than tests: syntax checks, unit checks on translation parity and verdict logic, a live smoke test, and a real browser pass in both themes at phone width

Tech stack

Node.js · TypeSafe AI Jev · Progressive Web App · Service Worker · Web Speech API · Docker · Dokploy · Traefik · GitHub Actions · Puppeteer

Project info

Year

2026

Duration

2026

Team

Solo

Role

Designer and Software Engineer

Status

Live

Category

Public Safety