Sajag (सजग, “alert”) is a Nepali-first scam safety companion, powered by Gemma 4. It checks a seller’s chat, payment QR or voice note in the few seconds before you pay, and if you’ve already paid, walks you through recovery and fills in the Cyber Bureau application.

Online fraud is the largest category of cybercrime complaints in Nepal, and once money leaves an eSewa or Khalti wallet it is very hard to get back. The only safe moment is before the QR is scanned.

What it does

• Check before paying: share screenshots, a payment QR, a voice note or pasted text (Nepali, Romanized Nepali or English). Gemma 4 reads it and streams what it finds live.
• Scam X-Ray: every warning is marked on the user’s own screenshot, with the exact line that triggered it.
• QR payee check by code, not AI: decodes the EMVCo payment QR and warns when it pays someone other than the shop in the chat, or when the checksum looks tampered.
• Hacked-friend check: “Didi needs money and a code” → call her on the number you already have.
• Sajag Agent: Gemma 4 function calling + Google Search checks the shop/page and price and writes polite Nepali messages to test the seller (it refuses to search private data).
• Already paid: “Did it arrive?” check-ins on the promised date, a recovery plan in Nepal Rastra Bank’s order with verified eSewa, Khalti, NRB and Cyber Bureau contacts, an OTP-emergency screen (incl. eSewa’s 48-hour insurance window), and the Cyber Bureau “Online Financial Fraud” application drafted by Gemma in Nepali with a Bikram Sambat date and fingerprinted (hashed) evidence annexes, ready to print and sign.
• Bank SMS reader by code: amount, debit/credit, transaction ID and date.
• Accessibility and privacy: read-aloud in Nepali, “hide details” blacks out OTPs/balances/faces before upload, cases stay on the phone with no account.
• On-device Gemma 4 E2B (LiteRT-LM): the check runs fully offline on the phone’s GPU (59 s on a mid-range Galaxy A55), so evidence never leaves the phone.

Why judges can trust the output

Gemma reads and points; code verifies and decides. Every warning must quote the evidence word-for-word or it’s dropped; the risk level is computed by code from community-written pattern files (patterns/*.yaml, each with a “not always a scam” note). Sajag shows warning signs, not verdicts, and never submits anything on the user’s behalf.

Open source

MIT-licensed. Scam patterns and verified contacts are plain YAML that anyone can improve with a PR. Includes an eval suite (28 end-to-end cases incl. honest sellers, all passing), 40 backend tests, and an agent skill (skills/sajag-check/SKILL.md) that lets any coding agent run a Sajag check.

Apps: Flutter for Android and iPhone, plus a web app. Team Markovian: Viraj Sawad, Aakriti Pandey, Chandan Kumar Shah.