$ cat projects/sanket-scam-shield.md
Sanket — Real-Time Scam Detection Copilot
Forward a suspicious message, get a plain-language verdict in your own language.
// why: First-time internet users had no fast way to check a message before acting.
TL;DR
- Forward anything: message or screenshot → scam / suspicious / safe verdict
- Always answers: 12 offline signatures first, then 4-provider LLM failover
- Vernacular + action: plain-language reply with one-tap 1930 escalation
// what does it do?
Anti-scam copilot for India over WhatsApp, SMS, web, and a browser extension. Forward a message or screenshot, get a scam / suspicious / safe verdict in plain vernacular language, with one-tap 1930 escalation.
// why this project?
India loses thousands of crores to digital scams every year. The victims are often first-time internet users, and they have no fast way to check a suspicious message before acting on it.
// what broke / learned
Offline-first was non-negotiable — the rule engine had to give a verdict even with zero API keys.
// results
- There's always an answer, whatever the infrastructure looks like. An offline rule pre-scan (12 India-specific scam signatures plus URL heuristics) runs first, then an LLM layer refines the result, with Claude fronting Gemini, Groq, and OpenRouter free-tier failover. If no API keys are configured at all, a rule-based fallback still produces a verdict. It also analyzes screenshots through vision models, replies natively in vernacular languages, generates shareable scam reports, publishes anonymized stats, and verifies webhook calls with HMAC.
Tech stack