Familieværn

Overview

• TwinCurrent built Familieværn, an AI proof-of-concept, at RoyalHacks 2026, Denmark's national student hackathon. It is an intelligent call-screening system designed to protect elderly Danes from phone scams: a complete voice pipeline, built end to end, where incoming calls are answered by an AI that listens, understands, and decides whether a call is a likely scam before it ever reaches the person being protected.

• The context is stark: phone fraud costs Danes large sums every year, and voice scams still get past existing filters. That is exactly the gap Familieværn targets.

The Challenge

• Voice scams still get past existing filters

• The 2024 government package closed SMS scams but left the voice channel open

• Elderly people are the most exposed and the hardest to reach with text-based defences

• Detecting a scam requires understanding spoken Danish in real time, not keyword matching

• Any screening must be fast enough to handle a live phone call without awkward delays

System Architecture

• A call arrives through Sinch telephony and is handed to a Rust/Axum backend on Railway that orchestrates the whole loop. Audio is transcribed by ElevenLabs Scribe, the transcript is classified by a Gemini Flash model in JSON mode, and the spoken response is synthesised by ElevenLabs Flash in a Danish voice.

• The backend streams every step (transcript, classification, decision) to a Next.js operator dashboard over Server-Sent Events, so a human can watch calls unfold live. The SSE pattern mirrors TwinCurrent's existing real-time chat plumbing, which let the whole pipeline come together inside a hackathon timebox.

The Solution

• Inbound call answering

  Calls are answered automatically through Sinch telephony, so the protected person's phone never has to ring for an unknown caller.

• Real-time transcription

  ElevenLabs Scribe turns the caller's spoken Danish into text the system can reason over as the conversation happens.

• Scam classifier

  A Gemini Flash model in JSON mode classifies the transcript, scoring how likely the call is a scam.

• Danish voice responses

  ElevenLabs Flash synthesises natural Danish speech so the AI can interact with the caller.

• Live operator dashboard

  A Next.js dashboard streams transcripts and classifications live over SSE so a human can monitor calls.

• Voice channel coverage

  Targets the scam channel left open after SMS fraud was clamped down, designed to protect the most exposed group.

Integration Highlights

• Sinch telephony

  Receives the inbound call and bridges audio to the backend.

• Gemini Flash + ElevenLabs

  Classifier in JSON mode plus Scribe STT and Flash Danish TTS.

• SSE operator stream

  Backend streams live events to the Next.js dashboard.

Tech Stack

• Rust / Axum

  Pipeline orchestration backend on Railway

• ElevenLabs Scribe + Flash

  Speech-to-text and Danish text-to-speech

• Gemini Flash

  Scam classifier in JSON mode

• Next.js + SSE

  Live operator dashboard on Vercel

Outcomes

• A complete inbound voice pipeline built end to end at RoyalHacks 2026

• Live Danish speech-to-text, scam classification, and text-to-speech in one loop

• Operator dashboard streams transcripts and decisions in real time over SSE

• Targets the voice scam gap left open after the 2024 SMS clampdown

• Built on a production-shaped stack (Rust/Axum, Railway, Vercel) rather than throwaway demo code

Have a real-time AI idea to prove out?

Familieværn went from concept to a working voice pipeline in a hackathon. If you have something to prototype fast, let's talk.

Have a similar project? Get an estimate within 24 hours See more cases