A pre-flight check for voice AI agents — built by an engineer who has shipped production voice agents on Bolna, Vapi, Retell, LiveKit, and ElevenLabs Conversational AI, and debugged the bugs that only show up at 2am on a live call. Paste in your voice agent's system prompt, opening line, and a one-line context (platform, use case, language). Get back a senior-engineer code review of your agent's configuration covering: - Race conditions in turn-taking — the "user-says-hello-during-the-welcome-message" bug that breaks most first deployments on Bolna and Vapi. - Prompt ambiguity and missing fallback paths — instructions the LLM can interpret two ways, undefined trigger conditions, no exit conditions, no escalation paths. - TTS-readability and latency issues — URLs and currency symbols read literally, monologues that bloat first-token latency, prompts that exceed practical token budgets for voice. - Safety, PII, and human-handoff gaps — missing consent disclosures, promises the agent structurally can't keep, no guardrails for hostile or distressed callers. Every issue comes with the exact phrase that's risky, why it matters in production, the user-experience impact, and a copy-paste-ready rewrite. Not generic advice. Best used on the day before you ship a new voice agent — or right after you fix a production incident, to catch the next one before it bites.