Content Filtering Voice Assistant with Guardrails

A voice AI assistant that answers questions about Cartesia, voice AI technology, and the competitive landscape while enforcing conversation guardrails. The agent preprocesses user inputs to detect and block toxic content, prompt injection attempts, and off-topic conversations, with escalation policies for repeated violations.

No demo yetView source Fork template

The numbers

latency—

cost / min—

frameworkCustom

The stack

telephonyWeb Only

speech-to-textCartesia STT

llmClaude Sonnet 4.5

text-to-speechCartesia Sonic-3

System prompt

No prompt published.

Config

config.json

{
  "temperature": 0,
  "tools": [
    {
      "name": "web_search",
      "description": "Search the web for up-to-date information about Cartesia, voice AI, and competitive landscape"
    }
  ],
  "architecture": "two-llm-design",
  "enforce_topic": true,
  "allowed_topics": "Cartesia AI products and technology, Voice AI/TTS/speech synthesis, AI/ML and software engineering, Competitors (ElevenLabs, PlayHT, Amazon Polly, etc.), Voice AI market and landscape",
  "block_toxicity": true,
  "guardrail_model": "gemini-2.5-flash-preview-09-2025",
  "inner_agent_model": "claude-sonnet-4.5",
  "guardrail_llm_model": "gemini-2.5-flash",
  "block_prompt_injection": true,
  "max_violations_before_end_call": 3
}

Tools

web_searchSearch the web for up-to-date information about Cartesia, voice AI, and competitive landscape

Voice AI recipes, picks, and analysis.

Get the useful new templates plus the occasional teardown of what’s working in production voice AI.