Voice agents that book the table, not the meeting after.
Sub-second turn latency. Multi-tenant by default. Production voice infrastructure for businesses that answer the phone: restaurants, clinics, salons, trades. Ship a branded agent the same day.
latency budget
Sub-second turn latency is what separates an agent that feels like a person from an agent that feels like a phone tree. Every component of the pipeline is streaming. Nothing waits for completion before passing to the next stage.
~450ms
sub-second turn latency
capabilities
Six things that decide whether a voice agent survives contact with real users on a real phone network.
Multi-tenant
Each tenant runs on a custom system prompt, dedicated phone number, industry preset, and isolated data scope. White-label ready: your logo, your domain, your billing surface. Supabase row-level security enforces the boundary at the data layer, not just the application.
Multi-provider LLM
Provider goes down, traffic reroutes in under 200ms. No conversation interruption. Track quality and latency per provider, route by tenant or by task.
7 industry presets
Hotels, restaurants, medical, dental, salons, and auto service. Each preset includes prompt, escalation rules, and tool inventory tuned for the vertical.
Live observability
Watch the call as it happens. See the LLM's reasoning, every tool call, sentiment trajectory. Take over mid-conversation if the agent gets stuck.
Tool execution
Google Calendar, Outlook, and Calendly sync. FAQ retrieval over a per-tenant knowledge base. Smart routing using full conversation context. Voicemail transcription with sentiment.
Phone-tuned audio
Deepgram Nova-3 streaming STT tuned for phone codecs. Cartesia Sonic-3 streaming TTS with ~40ms first-byte. LiveKit handles SIP trunking and barge-in detection. VAD calibrated for noisy environments. The audio path that decides whether the call sounds like a person or a phone tree.
architecture
An edge API orchestrates the call. A Python worker runs the voice pipeline on LiveKit. A multi-tenant dashboard observes everything in real time.
caller
PSTN · SIP trunk
livekit
WebRTC · barge-in · trunking
tenant
prompt · number · industry
Python · LiveKit worker
Hono · TypeScript
Next.js 15 · React 19
Deepgram Nova-3
streaming STT · ~150ms
OpenAI tool-calling
LLM · ~200ms TTFT
Cartesia Sonic-3
streaming TTS · ~40ms
industry presets
Each preset includes a domain-tuned prompt, escalation rules, tool inventory, and a baseline knowledge base. Pick a preset, set a phone number, ship.
stack
Deepgram Nova-3
Streaming STT
Cartesia Sonic-3
Streaming TTS
Gemini / OpenAI / Groq
LLM with fallback
LiveKit
SIP + WebRTC
Hono
Edge API
Next.js 15
Dashboard
Python
Agent worker
Supabase
DB + realtime
talk shop
Live call monitoring, tenant configuration, deal pipeline, sentiment tracking: the whole supervisor surface. Send a note and I'll walk you through it.