Nandi: Autonomous Voice Sentinel & Scam Guardian

Real-time edge AI guardian against digital arrest and extortion scams, powered by NVIDIA Nemotron reasoning, Gnani vernacular voice, and Tavily.

Comment

Inspiration

Across India, thousands of vulnerable citizens, especially retirees, students, homemakers, and technologically less-informed individuals, are falling victim to sophisticated "Digital Arrest" and authority impersonation scams. Fraudsters posing as officers from the Central Bureau of Investigation (CBI), Narcotics Control Bureau (NCB), Telecom Regulatory Authority of India (TRAI), or state police forces coerce victims over video or phone calls. By fabricating arrest warrants, citing seized parcels with narcotics, and weaponizing legal terminology, they isolate victims in their rooms and force them into liquidating their life savings into fake "escrow verification" accounts before family members or banks notice.

According to the Ministry of Home Affairs (MHA) Indian Cyber Crime Coordination Centre (I4C), victims lost over ₹1,750 Crore to digital arrest scams in recent quarters alone.

Our inspiration comes from Nandi, the mythical divine bull in Indian tradition: the steadfast threshold guardian (Dvarapala), embodying attentive hearing (Sravana), unwavering stillness (Achala), and absolute loyalty. We asked ourselves: Can we build an ambient, always-on AI guardian that sits quietly at the household threshold, listens protectively, cuts through synthetic intimidation with cold legal reality, and intervenes before money moves?

What It Does

Nandi is a real-time, privacy-preserving edge AI guardian that detects, verifies, and halts digital arrest and authority extortion scams as they happen.

  • Autonomous Threat Tracking: Implements a deterministic 7-stage scam state machine (Hook $\rightarrow$ Authority Impersonation $\rightarrow$ Threat Escalation $\rightarrow$ Isolation Channel $\rightarrow$ Confinement $\rightarrow$ Fake KYC $\rightarrow$ Asset Liquidation) running with sub-millisecond evaluation latency ($0.088\text{ ms}$ median).
  • NVIDIA Nemotron Reasoning: Analyzes conversational context through Nebius Token Factory to extract officer names, badge numbers, forged court orders, and implicit intimidation cues.
  • Pan-India Cyber Jurisdiction Verification: When callers claim jurisdiction, Nandi uses edge geolocation and Tavily search scoped strictly to official Indian domains (.gov.in, .nic.in, rbi.org.in, npci.org.in) across all 28 States and 8 Union Territories. It exposes fake notices against I4C Advisory TAU/ADV/003 (confirming that no government agency conducts digital arrests or demands money transfers on calls).
  • Immediate Vernacular Voice Interventions: If risk crosses into DANGER before payment demands execute, Nandi triggers an audible in-room intervention using Gnani Timbre v2.5 across four selectable personas (Deepak, Kaveri, Radhika, Pranav) in Hindi and Hinglish: > "तुरंत रुकिए! कोई भी पुलिस या सीबीआई फोन पर डिजिटल अरेस्ट नहीं करती। किसी को पैसे ट्रांसफर न करें।"
  • Family Safety Console & Streamable MCP: Family members receive instant escalation alerts and can inspect full evidence trails via our React console or connect through the Model Context Protocol (Streamable HTTP MCP) to file pre-populated cybercrime reports to Chakshu and National Cyber Crime Reporting Portal (1930).
  • DPDP Act 2023 Compliance: Zero raw audio is ever stored or uploaded. Audio is transcribed in volatile memory, converted to redacted forensic cues, and dropped immediately. 30-day automated TTL purge is enforced on all metadata.

How We Built It

We designed Nandi as a hybrid edge-and-cloud architecture:

  1. Reasoning Engine (Nebius Token Factory & NVIDIA Nemotron): We utilize nvidia/Nemotron-3_5-Lightning on Nebius Token Factory for fast, high-accuracy conversational extraction, with Nemotron 3 Ultra handling complex multi-party forensic analysis. Token Factory provided serverless OpenAI-compatible inference with ultra-low token generation latencies.

  2. Vernacular Speech Layer (Gnani AI):

    • Prisma v2.5 STT: Real-time streaming transcription (wss://api.vachana.ai/stt/v3/stream) supporting conversational Hindi, Indian English, and code-mixed Hinglish.
    • Timbre v2.5 TTS: Low-latency neural synthesis (https://api.vachana.ai/api/v1/tts/inference) delivering 48 kHz PCM16 audio, packaged into native WAV containers at the edge across 4 distinct voice personas.

  3. Ground Truth Verification (Tavily Search): Edge geolocation detects the user's nearest state or union territory. Nandi then dispatches regional search queries to Tavily, filtering out SEO spam and cross-referencing claims against local police cyber cells, RBI notification circulars, and Supreme Court rulings.

  4. Edge Deployment (Cloudflare Workers): The entire system runs 24/7 globally on Cloudflare Workers edge runtime at https://nandi.erebuzzz.tech, with zero cold starts and distributed state management via stateless URL-safe cryptographic tokens.

  5. Client Architecture & Design Genome (Pixasso System): Built with React 19, TypeScript, and Vite. Designed under the Pixasso system with liquid glassmorphism, 60fps live canvas seismograph waveforms, sacred geometry vector iconography, and dual AMOLED/Light theme engines.

Challenges We Ran Into

  • Audio Streaming & Containerization at the Edge: Gnani Timbre returns raw 48 kHz 16-bit PCM streams without RIFF headers. Browsers cannot natively decode raw byte streams without an AudioContext decoding pipeline. We engineered a zero-copy PCM-to-WAV packaging function directly in the Cloudflare Worker edge environment to dynamically compute byte rates and attach valid 44-byte WAV headers on the fly.
  • Deterministic Stage Progression vs. LLM Latency: Pure LLM evaluation on every audio chunk is too slow for real-time scam intervention. If a scammer demands an immediate RTGS or UPI transfer, waiting 3 seconds for an LLM response could mean a victim loses their life savings. We resolved this by building a deterministic, sub-millisecond stage tracker for immediate hazard protection, delegating asynchronous semantic claim extraction to NVIDIA Nemotron in parallel.
  • Zero-Knowledge Audio Privacy: Building an ambient room guardian without turning it into a privacy liability was paramount. We architected the system so that raw audio is strictly kept in volatile memory buffers and destroyed immediately upon transcription. Only anonymized, rule-extracted cues with redacted PII are stored in local SQLite containers, strictly following India's Digital Personal Data Protection (DPDP) Act 2023.

Accomplishments That We're Proud Of

  • 83.3% Alert-Before-Ask Benchmark: On our 12-scenario evaluation suite spanning code-mixed Hinglish, noisy room acoustics, and aggressive impersonation scripts, Nandi intervenes before the scammer's first payment or remote-access demand in over 83% of cases, with zero false alarms per hour on benign household chatter.
  • Full MCP Streamable HTTP Implementation: First scam-prevention system to implement the Streamable HTTP Model Context Protocol specification, enabling AI assistants and family dashboards to query live call risk and prepare Chakshu 1930 complaint drafts seamlessly.
  • Complete Pan-India Scoping: Dynamic territorial dispatch covering all 28 States and 8 Union Territories, ensuring rural and urban households alike receive regional cyber cell contact details.
  • Always-On Edge Production: Deployed live on a custom domain with global edge routing, sub-second vernacular voice synthesis, and zero server management overhead.

What We Learned

  • Acoustic Nuances in Vulnerable Households: Scammers deliberately modulate tone between authoritative shouting and fake legal politeness. Detecting this requires analyzing semantic intent and structural stage progression rather than audio volume alone.
  • High Efficiency with Nemotron on Nebius: NVIDIA Nemotron served on Nebius Token Factory delivered consistent generation speeds that made complex conversational classification viable in an asynchronous pipeline without GPU warm-up delays.
  • The Importance of Human-Centric UI: When vulnerable users are panicking, dense technical alerts fail. Showing clear color-coded states (Safe, Watch, Warn, Danger), audio chime confirmations, and an immediate actionable message in their native language is what actually prevents harm.

What's Next for Nandi

  • Hardware Acoustic Pod (Nandi Karani): Developing a dedicated low-cost ambient desk hardware puck with directional microphone arrays and local edge inference for households without an active browser session.
  • Telco-Grade SIP/VoIP Gateway: Collaborating with telecom providers to deploy Nandi directly at the carrier gateway level to filter incoming spoofed WhatsApp and international VoIP calls before they ring.
  • Expansion to 12 Indian Scheduled Languages: Expanding Gnani Prisma and Timbre integration from Hindi and Hinglish to Tamil, Telugu, Bengali, Marathi, Kannada, Gujarati, and Punjabi.

Built With

  • ai-agents
  • cloudflare-workers
  • cyber-security
  • digital-arrest
  • dpdp-act
  • edge-computing
  • fastapi
  • gnani-ai
  • model-context-protocol
  • nebius-token-factory
  • nvidia-nemotron
  • pan-india
  • privacy
  • python
  • react
  • scam-detection
  • speech-to-text
  • sqlite
  • tailwindcss
  • tavily
  • text-to-speech
  • typescript
  • vite
  • web-audio-api
  • web-speech-api
Share this project:

Updates

Submission history