TITANS

AppShield AI: a real-time guardian that stops fraud before it happens The numbers are real and growing. Indians lost ₹22,845 crore to cyber fraud in 2024, a 206% rise over 2023, according to Government of India data. Behind each figure is a family. An elderly father receives a WhatsApp video call from a “CBI officer”, is told his Aadhaar is linked to money laundering, and is kept in “digital custody” for hours. He is warned not to tell anyone and pressured into transferring his life savings. By the time his family finds out, the money is gone. Blocking numbers doesn’t help, because scammers switch SIMs every day. What never changes is the manipulation inside the conversation. AppShield uses Meta’s AI stack to stop the fraud before it happens. It acts as a guardian on the call, recognising the scam as it builds and stepping in before the victim shares an OTP, approves a UPI request, installs a remote-access app or sends money. How it works. With the user’s consent, the conversation is transcribed live with Muse Voice Transcribe. Every line then goes through four steps: AppShield-Tagger, our fine-tune of Meta’s XLM-RoBERTa, detects 18 manipulation tactics across English, Hinglish and Hindi, including fake authority, threats, urgency, isolation, secrecy and identity harvesting. ScamStage-Mini tracks the scam’s progression and predicts the chance of a harmful request in the next 3–5 turns. A risk score is calculated for the conversation. A clear action is chosen: Monitor, Caution, Warn, Verify, Stop or Guardian. Muse Glimmer turns the evidence into a warning a frightened person can trust: “They claimed to be CBI, told you to keep it secret, and are now asking for money. Real police never ask for transfers. Hang up.” Three layers of protection: Real-time call detection. The warning comes before the harmful ask, not after the loss. Guardian alert. With the user’s confirmation, a family member is alerted instantly, with a summary from Muse Wonder, the risk timeline and what to do next. Safe Word Rescue. If the victim is trapped on the call, saying a pre-agreed safe word silently notifies their guardian. SAM 3.1 blurs QR codes and faces in screenshots the user shares, protecting privacy. Built to be trusted. AppShield never takes control from the user. No raw audio is stored, transcripts are deleted on the user’s chosen schedule, and guardians see only what the user allows. Stack: FastAPI on Cloud Run, Firebase, a Kotlin Android app, a guardian web dashboard, and models trained on an H100.