[Suggestion] Voice agent with conversational guardrails: topic enforcement, PII blocking, and response validation (Python)

[deepgram-robot]

What to build

A recipe demonstrating how to add safety guardrails to a Deepgram voice agent — enforcing topic boundaries, blocking PII in agent responses, validating outputs against business rules, and gracefully redirecting off-topic conversations back to the intended scope.

Why this matters

Developers deploying voice agents in regulated industries (healthcare, finance, legal) need guardrails that prevent agents from discussing unauthorized topics, leaking sensitive information, or providing advice outside their scope. Deepgram's Audio Intelligence features (PII redaction, topic detection, intent classification) combined with the Voice Agent API's function calling provide the building blocks, but no reference implementation shows how to wire them together into a comprehensive guardrails layer. This is a blocking requirement for enterprise voice agent deployments.

Suggested scope

• Language: Python
• Deepgram APIs: Voice Agent API (WebSocket), STT with Audio Intelligence (PII redaction, topic detection, intent classification), TTS (Aura)
• Guardrail types:
  1. Topic boundary enforcement: Detect off-topic intents and redirect
  2. PII blocking: Redact PII from both user input transcription and agent responses before TTS
  3. Response validation: Check agent output against configurable business rules before speaking
  4. Escalation trigger: Detect anger/frustration sentiment and offer human handoff
• Includes: Configurable guardrails rules (YAML/JSON), logging of blocked/redirected interactions, metrics on guardrail trigger frequency
• Complexity: Advanced

Acceptance criteria

• Runnable with minimal setup (clone, add API key, run)
• README explains each guardrail type with examples of what gets blocked/redirected
• Uses current SDK version
• Configurable guardrails via YAML or JSON rules file
• Demonstrates PII redaction in both directions (user → agent, agent → user)
• Logs all guardrail interventions for compliance auditing
• Includes test conversation scripts that trigger each guardrail type

---

Raised by the DX intelligence system.

---

Queued by PM — Engineer will pick this up as a priority:user recipe.

[claude]

@deepgram/devrel — VP escalation: Issue #565 has been open >4 hours with no bot response. Attempted to re-trigger pm-suggestions workflow but GITHUB_TOKEN lacks dispatch permissions. Manual triage needed.

State: No labels, no bot comments, no queue labels applied.

[github-actions]

Ooo, we'll get right on that! We'll build a voice agent guardrails recipe for you — demonstrating how to use prompt-based topic enforcement and function calling for escalation in a Deepgram voice agent.

A quick note on scope: recipes in this repo are focused, runnable examples (<50 lines), so the Engineer will distill the core guardrails pattern (topic boundary enforcement via the system prompt + an escalation function call) into a clean, minimal example. Features like PII redaction in STT already have their own recipe path and can be composed alongside this one.

This has been queued for the Engineer — it'll get priority as a user request.

[claude]

VP check-in (2026-05-25 16:58 UTC): This queue issue has been open >6 hours without being picked up by the engineer workflow. Attempted to re-trigger engineer.yml but workflow dispatch is not permitted with the current token. The next scheduled engineer cron run should pick this up automatically.

[github-actions]

Opened 1 PR — #566 (voice-agents v1 guardrails recipe for Python). Waiting for human review and merge.