AURALERT — Voice-First Scam Defense

AURALERT is a voice-first scam defense tool that listens to suspicious calls, checks family safe words, links repeat-offender indicators, and prepares incident reports.

Comment

Inspiration

Cyber fraud, social engineering, and extortion syndicates cause losses estimated at roughly $1 trillion a year worldwide (Global Anti-Scam Alliance, Global State of Scams 2024). In emerging markets, scams such as "Digital Arrest" (fake CBI/police video calls), UPI Collect Deceptions, and Fake KYC Expirations inflict immense emotional and financial trauma on victims, particularly seniors.

When a threatening call or fake text arrives, victims panic. They do not know how to inspect lookalike URLs, verify police credentials, or find emergency helpline numbers. We asked: What if a smart voice assistant could be the calm, defensive second opinion in the room?

What it does

AURALERT is a voice-first scam defense tool built for Alexa+ that listens to suspicious calls, checks family safe words, links repeat-offender indicators, and prepares incident reports.

  1. Live Call Guard on Speakerphone: Put a suspicious call on speakerphone. AURALERT listens in rolling chunks, tracks cumulative risk across the conversation, and interrupts with spoken guidance ("They just asked for an OTP, hang up immediately!").
  2. Voice Triage & Red-Flag Breakdown: Speak a suspicious call or message aloud and get a clear verdict in at most two calm spoken sentences, always leading with the safest immediate action.
  3. Family Safe-Word Check: Compares what a caller claims against a privately configured family passphrase, helping defend against voice-cloning and fake-emergency scams.
  4. Repeat-Offender Memory: Matches phone numbers, UPI handles, and domains against previously reported indicators in a local database, with exportable indicator reports.
  5. Static Link Inspection: Audits lookalike domains, Punycode, raw IP hosts, shorteners, and high-risk TLDs. Links are never fetched, so the tool cannot be tricked into contacting attacker-controlled servers.
  6. Incident Report Generator: Compiles a structured report with case ID, timeline, and official reporting channels (India 1930 and cybercrime.gov.in, US FTC/IC3, UK Report Fraud).
  7. Elder Scam Drills & Hindi Voice Mode: Practice drills that score how well a user resists pressure, with voice support in English and Hindi.

How we built it

  • Model Context Protocol (MCP) Server: Built with the official Python MCP SDK, serving 17 tools over Streamable HTTP at /mcp with Origin and Host validation, on a Starlette application with rate limiting.
  • Deterministic Defensive Engine: Explainable, rule-based heuristics across 11 scam categories in English and Hinglish, with benign-notice handling (legitimate bank messages that say "do not share your OTP" are not flagged in our tests). Detection is rule-based, not a trained model.
  • Agent Brain: A deterministic local router that works without cloud access, plus an optional OpenRouter brain and an optional Amazon Bedrock Converse integration (implemented, but not used in the live demo). Every response reports which brain answered.
  • Simulated Alexa+ Interface: Built with the browser's Web Speech API for voice in and out, with a risk meter, case-file panel, and visible MCP tool-call chips. It is a simulation, not an official Alexa client.
  • Testing & Quality: 81 automated tests covering detection logic, MCP transport security, agent routing, and adversarial inputs such as prompt injection.

Challenges we ran into

  • Minimizing False Alarms: Real bank messages use the same words as scams ("OTP", "immediate action"). Our first Live Call Guard even told a user to hang up on a genuine bank warning, so we added rules that ignore protective phrasing and only score the caller's words.
  • Voice-First Constraint: Spoken output has to stay under two sentences and lead with the safest action, while detailed evidence appears on screen.
  • Safe Link Analysis: Analyzing suspicious links without ever visiting them.

Accomplishments that we're proud of

  • 81 automated tests, all passing, including adversarial and security cases.
  • A complete flow from live speakerphone listening to a complaint-ready incident report in a single session.
  • English and Hindi voice support.

What we learned

Defensive voice AI requires restraint: tell the user the one safest action first, and keep the forensic detail on screen. We also learned to test our own tool against normal calls, not just scams.

What's next for AURALERT — Voice-First Scam Defense

  • Running on physical Alexa+ devices.
  • Real, consent-based family notifications (today the family alert is only prepared, not sent).
  • Direct filing into national cybercrime reporting portals.
  • More regional languages and dialects.

Built With

Share this project:

Updates

Submission history