Pipeline, realtime, or self-hosted — with tools, escalation and evals built in.
Any ASR, any LLM, any TTS, wired as a DAG. Each agent’s config lives in the portal and hydrates at call time — swap a provider without a deploy.
OpenAI Realtime and Grok realtime with per-agent server-VAD tuning — or Qwen Omni self-hosted on your GPUs via vLLM, replacing the whole pipeline with one model.
Turn detection plus energy VAD live in the platform, not in a provider SDK. The instant a caller speaks, the agent stops — identically on every model.
route_to_skill hands the call to a live human queue. transfer_call moves it anywhere — blind, warm or consult. end_chat closes it. present_form collects structured input. HTTP and MCP tools reach everything else.
AI tool call → skill queue → hold music → claim → ring → bridge to a human. Live end to end on PSTN, not a demo.
Simulated callers — scripted or LLM personas — hit your agent with barge-in timing checks and function-call expectations. LLM judges grade transcripts; suites roll up p50/p95 first-audio. Runs hermetically in CI with zero keys.
Agent Assist taps live call audio into a tenant-keyed ASR+LLM: live transcript, suggestion cards, and an AI wrap-up draft in the agent’s workspace.
Telequick runs every layer of voice — you bring the models.