ScamLens

Problem

KYC-expiry fraud, UPI collect-request tricks, "digital arrest" intimidation calls, and fake work-from-home job offers target millions of people in India every year — often in a mix of English, Hinglish, and Tamil. Most victims never get a second opinion before they act.

What it does

ScamLens is a scam-message screening assistant. Paste a suspicious SMS, WhatsApp message, email, or call transcript and it returns a SAFE / SUSPICIOUS / SCAM verdict plus confidence, quotes the exact red-flag phrases (urgency, authority impersonation, OTP/payment/UPI requests, suspicious links, too-good-to-be-true offers), explains the trick in plain English AND Tamil, and gives an action checklist (verify through official channels, never share OTPs/PINs/CVV, report fraud at cybercrime.gov.in, helpline 1930).

How it works

1. Deterministic pre-screener — regex artifact extraction (URLs incl. URL shorteners, UPI IDs, phone numbers, OTP-like digit runs) plus multilingual keyword lexicons (English/Hinglish/Tamil transliteration).
2. LLM verdict — NVIDIA Nemotron 3 Super returns strict JSON: verdict, confidence, verbatim-quoted red flags, plain-English trick explanation, safety checklist.
3. Weighted rules-based fallback if the LLM fails or returns garbage ("mode": "llm" vs "mode": "rules"), so the tool never goes silent.
4. The Tamil explanation is a deterministic human-written template matched to the verdict — the model cannot reliably generate Tamil-script prose, so it is not asked to.

Stack

FastAPI + SQLite (stdlib sqlite3) + vanilla JS; NVIDIA Nemotron 3 Super via the OpenAI-compatible chat-completions API; pytest — 69 tests passing.

Honesty notes

• The six sample_inputs files are self-authored archetypes of publicly documented scam patterns (plus two genuine messages) — illustrative fixtures, not real messages; phone numbers, UPI IDs, and links in them are fictional.
• Verdicts are advisory, not legal or financial advice.
• No accuracy claims — the tool has not been evaluated on a labeled dataset.
• Built entirely during the ForgeHacks event window (Oct 3–10, 2026).

Built With

• ai
• cybersecurity
• fastapi
• fraud-detection
• javascript
• llm
• nemotron
• nvidia
• python
• sqlite