ScamShield — AI Scam & Link Verifier

The Problem: Scam messages exploit urgency and impersonation to steal money and data — and in India, this isn't a fringe issue, it's a crisis. Indians lost ₹19,812 crore to cyber fraud in 2025 alone, with Maharashtra hit hardest at ₹3,203 crore (I4C data). The messages driving this are getting more convincing every year: a real example that inspired this project is an SMS impersonating MSEDCL (a state electricity board) demanding payment of a traffic fine, linking to "mparivahanofficial.com" — a near-perfect lookalike of India's real government portal, parivahan.gov.in. Most people have seconds to decide whether to trust a message like this, and existing spam filters only scan for keywords — they never check whether the link itself is real. The Solution: ScamShield closes that gap by running two independent, verifiable investigations on every message instead of a single AI guess. Content Analysis uses an LLM to detect manipulation patterns: urgency, OTP requests, mismatched sender identity, financial threats. Link Forensics checks the URL against real, objective data an LLM cannot fabricate — domain age via WHOIS, SSL certificate issue date, hosting/ASN location, and cross-referenced reputation via Google Safe Browsing and VirusTotal. Emails additionally get SPF/DKIM/DMARC authentication checks. For SMS, where sender ID is trivially spoofed, we say so directly instead of pretending to verify the unverifiable — a deliberate honesty choice that makes the tool trustworthy rather than overconfident. Every check feeds into one explainable verdict (HIGH/MEDIUM/LOW risk) with every contributing reason listed in full. Target Users: individuals and families receiving suspicious SMS/email — this is a working, complete tool for them today. As a cybersecurity and digital forensics student who has interned with Maharashtra Cyber, I've seen these exact fraud patterns firsthand, and that direct field experience shaped every check this tool performs. How They Interact: paste a message (or several, in batch mode), select SMS or email, and get a structured investigation report — Sender Analysis, Content Analysis, Link Forensics, Final Verdict — built to look and read like an actual security analyst's findings, not a spam-filter popup. What Makes It Unique: nearly every scam detector on the market is either a keyword filter or a pure LLM classifier — both of which can be fooled, and the latter can hallucinate about whether a link is even real. ScamShield is architected around a reasoning layer paired with a facts layer, making it explainable and defensible in a way black-box classifiers simply aren't. Roadmap: WhatsApp Business API integration for inline protection, a browser extension, and a Granite-fine-tuned model trained specifically on regional Hinglish/Indian scam patterns. Longer-term, the underlying detection logic could be offered as an API that bank or telecom fraud teams integrate into their own review workflows to triage suspicious messages faster.