RakshaAI rides alongside a call, message, or UPI request in real time — detecting digital-arrest scripts, cloned voices, and social-engineering pressure the moment they start, and pulling in a trusted family member before money moves.
This is a real, working detector — pick a call scenario and watch RakshaAI score risk turn-by-turn, exactly as it would on a live phone call. Or paste any suspicious message into the scanner.
A pure LLM call is too slow and too fragile for a live phone call. RakshaAI pairs an instant offline pattern engine with an LLM reasoning layer for semantic, never-seen-before scripts.
Live call audio (via telecom/OEM accessibility APIs) or pasted SMS/WhatsApp text is transcribed on-device in near real time.
A <50ms offline pattern engine scores known scam signatures — impersonation phrases, urgency, secrecy, OTP/fund-transfer requests — even with no network.
Claude analyzes full conversational context and intent to catch novel, paraphrased, or multilingual scam scripts the heuristic engine misses.
Above threshold: on-screen warning in the caller's language, call-recording prompt, and one-tap alert to a pre-set family guardian with live risk context.
Not a hypothetical — ₹22,495 crore lost in 2025 alone, with digital-arrest scams growing fastest. Every judge's parent or grandparent is the target user.
Scammers now use voice cloning and AI-generated scripts. RakshaAI is explicitly AI-vs-AI — a sharper, more defensible narrative than "AI for X."
The offline heuristic layer keeps protecting users through poor rural connectivity — critical since first-time digital users in Tier-2/3 India are now the fastest-growing target group.
Licensable to banks (fraud-prevention mandate under RBI), telecoms (caller-risk layer), and insurers — not just a consumer app with no business model.
Elderly and first-time digital users, not developers — large text, voice-first alerts, regional language support, and a family safety net instead of relying on the victim alone.
The detection core, dashboard, and scanner you're using right now are fully working — not a mockup. Judges can test it live.