Inspiration

Scams are evolving faster than awareness. Americans lost $12.5 billion to fraud in 2024 alone, and with AI-generated voices, deepfake video calls, and hyper-personalized phishing, traditional advice like "just don't click the link" is no longer enough. Elderly and vulnerable people are hit hardest, yet most don't have the tools to fight back. Unhooked was born from a simple idea: what if your phone could spot a scam before you fall for it, using the same AI that scammers are weaponizing?

What it does

Unhooked is an AI-powered scam detection app with three ways to protect yourself. Voice Agent lets you talk to a real-time AI agent. Describe a suspicious call, email, or situation and get conversational scam guidance on the spot. Text Scanner lets you paste any suspicious message, email, DM, or text and get an instant risk assessment with a 0-100 score, red flags, and actionable recommendations. Image Scanner lets you upload a screenshot of a suspicious website, message, or social media post, and Claude's vision AI reads and analyzes the content for scam patterns. Every analysis returns a clear verdict: Safe, Suspicious, or Likely Scam, so you know exactly what to do.

How we built it

Unhooked is a cross-platform Flutter app with a dark, security-focused UI. The voice agent is powered by ElevenLabs' Conversational AI SDK, connecting via WebRTC through LiveKit for low-latency real-time conversation. Text and image analysis are handled by Anthropic's Claude Sonnet 4 API. We crafted a system prompt that instructs Claude to return structured JSON with a verdict, risk score, red flags, and recommendations. For image scanning, we send base64-encoded images to Claude's vision endpoint so it can read and reason about screenshot content directly, no OCR pipeline needed. Everything runs client-side in a single consolidated Dart file.

Challenges we ran into

Getting the ElevenLabs voice agent working in Flutter was our biggest challenge. The SDK uses WebRTC under the hood, and we had to debug authentication flows. The agent endpoint returns a conversation token, not a direct WebSocket URL, and getting that handshake right took real effort. Voice latency was another battle; we learned that TTS model choice (Turbo v2.5 vs Multilingual v2) and LLM model selection dramatically affect response times. We also had to build a filter to strip ElevenLabs expression tags like [sad] and [laughing] from agent responses before displaying them to users.

Accomplishments that we're proud of

We're proud that Unhooked works end-to-end across all three modes, from a user pasting a phishing email and getting a detailed risk breakdown, to uploading a screenshot that Claude reads and analyzes, to having a live voice conversation with an AI agent about a suspicious situation. Building a polished, cohesive app with real-time voice AI, vision analysis, and text analysis in a hackathon timeframe is something we're genuinely proud of. The UI came together well too: color-coded verdicts, animated risk meters, and a home screen that makes the app immediately intuitive.

What we learned

We learned how powerful it is to combine different AI modalities into a single protective tool. Claude's ability to analyze both text and images with the same structured output format made the architecture clean and consistent. We also gained deep experience with real-time voice AI. ElevenLabs' agent framework is impressive but has nuances around latency, turn-taking, and expression tags that aren't obvious from the docs. The biggest takeaway: the best defense against AI-powered scams is AI-powered protection.

What's next for Unhooked

We want to move API keys to a proper backend server for production security, add support for more languages to help non-native speakers, and build a history feature so users can review past scans. We'd also love to add real-time phone call monitoring that can flag suspicious calls as they happen, a community reporting system where users contribute to a shared scam database, and a browser extension that scans pages in real time. The scammers keep evolving, and so will Unhooked.

Built With

Share this project:

Updates