BantayScam: Forward the text, get the truth in Taglish

Problem Statement

Text scams are part of daily life in the Philippines. Fake GCash and Maya “account verification” links, fake DSWD ayuda and 4Ps payout messages, fake package delivery notices, and look-alike government websites keep tricking people, especially seniors, OFW families, and first-time e-wallet users. The SIM Registration Act (RA 11934) and the Anti-Financial Account Scamming Act (RA 12010) have made scamming harder, but scammers have moved to new numbers, messaging apps, and spoofed sender IDs. Victims usually don’t know who to ask, and reports to the CICC 1326 hotline are mostly made after the money is already gone. Agencies also don’t get a fast, structured picture of which scams are trending in which regions.

Proposed Solution

BantayScam runs on Messenger, Viber, Telegram, and a Smart SMS short code. A user forwards the suspicious message, pastes a link, or uploads a screenshot. The AI reads the text (with OCR for screenshots), expands shortened URLs, and checks domains against official government and bank domain lists, domain age, and known phishing feeds. It then returns a verdict (Scam / Suspicious / Likely Safe) with a short Taglish explanation of the red flags, for example: “Hindi nagpapadala ang DSWD ng link para mag-claim ng ayuda. Bago lang ginawa ang website na ito (3 days old).” It also tells the user what to do next, such as reporting to 1326, contacting the bank or GCash, or blocking the number. Each report is stripped of personal data and grouped into scam campaigns. These campaigns feed a dashboard that shows CICC, LGUs, and telcos the trending scams, the brands being impersonated, and heatmaps by region.

Target Market

• Primary users: Filipino mobile users, about 100M+ registered SIMs and 80M+ e-wallet accounts. The focus is seniors, OFW families, students, and 4Ps beneficiaries.
• B2G buyers: CICC, DICT, NTC, DSWD (to protect ayuda channels), PNP-ACG, and LGU public information offices. There are about 1,700 LGUs.
• B2B buyers: Banks, e-wallets, couriers, and retailers whose brands get impersonated. They could pay for brand-abuse alerts and takedown evidence packs.

How AI Is Used

• Multilingual LLM classifier and explainer: It understands Taglish, Bisaya, and scam slang, and writes explanations a lola can follow. Keyword rules can’t keep up with scam scripts that change every week, and they can’t explain why a message is a scam.
• Vision/OCR on screenshots: Many scams arrive as images or are forwarded as screenshots. Without OCR plus layout understanding, such as spotting a fake GCash logo or a fake receipt, these reports can’t be checked at all.
• URL and domain reasoning: Features such as domain age, typosquatting (e.g., “gcash-verify[.]ph”), and redirect chains are combined into a risk score.
• Embedding-based clustering: Thousands of reports are grouped into campaigns (“Fake LTO fine SMS – NCR, 412 reports in 24h”). Without this, the dashboard would just be noise.
• Anomaly detection: It flags sudden spikes in a scam type or region so agencies can issue early warnings.

PLDT / Smart / ePLDT / Rokid Angle

• Smart: It fits Smart’s existing anti-scam and spam-filtering efforts. Smart can offer a free short code for reporting, and campaign intelligence can feed back into network-level SMS filtering.
• ePLDT: Hosts the dashboard and report data on Philippine data centers and cloud, which keeps sensitive government data in-country. It can be sold to agencies as a managed B2G service.
• PLDT Enterprise: Gives PLDT a “brand protection” product to sell to banks, e-wallets, and couriers, with an upsell to enterprise SMS sender-ID verification.
• Rokid (optional): Barangay helpdesk staff wearing AR glasses could glance at a resident’s phone and get an instant scam verdict overlay.

Demo Scenario (3 minutes for judges)

1. Show a real-looking (sanitized) text: “DSWD AYUDA: Ikaw ay napili para sa ₱8,000. I-click: dswd-ayuda-claim[.]com.” Say: “My lola got this last week.”
2. Forward it to the BantayScam Messenger bot live. In about 5 seconds it replies in Taglish: verdict SCAM, three red flags (non-gov.ph domain, registered 2 days ago, DSWD never sends claim links), and next steps (report to 1326, don’t enter your MPIN).
3. Upload a screenshot of a fake GCash “account suspended” message. The bot reads it with OCR and catches the look-alike domain.
4. Paste a legitimate gov.ph advisory to show the bot gives an honest “Likely Safe” with caveats.
5. Switch to the CICC/LGU dashboard. A campaign card shows “Fake DSWD ayuda – 37 reports in 2 hours, mostly Region IV-A.” It also shows the impersonated-brands chart and a one-click public advisory draft for the LGU’s Facebook page.
6. Close with: “Every forward protects the next person.”

30-Hour Build Plan

Suggested stack: Python FastAPI, Gemini or GPT-class multimodal LLM, Tesseract or Cloud Vision OCR, sentence-transformers + HDBSCAN, Supabase/Postgres, Next.js + Leaflet dashboard, Telegram/Messenger Bot APIs. Public data / APIs: CICC and PhilCERT public advisories, DICT and BSP scam bulletins, PhishTank / OpenPhish / URLhaus feeds, WHOIS/RDAP, Google Safe Browsing API, the gov.ph domain list, Hugging Face SMS spam datasets (for baseline only).

Risks & Privacy Notes

• False negatives and positives: Never say “100% safe.” Always show confidence and advise checking through official hotlines. Run regular red-team tests.
• Data Privacy Act 2012: Forwarded messages may contain the user’s name, phone number, or account number. These are auto-redacted before storage and before anything is sent to the LLM. Collect only what is needed, set a retention limit (e.g., 90 days), and show a clear consent notice in the bot.
• Dashboard aggregation: It shows aggregates only (minimum group sizes) and cannot be used to re-identify reporters.
• Clause 24 / competition rules: The demo uses only synthetic or publicly published scam samples. No real personal data goes into AI tools.
• Abuse: Rate-limit the bot so scammers can’t use it to test which messages slip through. Don’t expose the full classifier rules.

AI Use Disclosure (draft)

BantayScam uses a multimodal LLM (e.g., Google Gemini) to classify messages and write Taglish explanations, OCR (Tesseract or Cloud Vision) to read screenshots, and sentence-transformer embeddings to cluster reports. The team used AI coding assistants (e.g., GitHub Copilot, Gemini) for boilerplate code and pitch copy. The datasets, prompts, verdict logic, and dashboard design were built and reviewed by the team, and every AI output in the demo was checked by a human.

Rubric Scoring