AI Voice Agents

Give your Divinci release a phone number. Callers dial an ordinary number and talk to the same grounded assistant you ship on the web — your knowledge base, your fine-tune, your guardrails — with streaming answers that start speaking in seconds. No app. No browser. No sign-up.

Talk to us Read the engineering deep-dive →

What happens when someone calls

A Voice Agent is not a separate chatbot with a microphone bolted on. The call rides the exact same release pipeline as your web chat — retrieval, fine-tune, fallbacks, moderation, billing — with a real-time voice loop wrapped around it.

caller dials your number → SIP trunk → LiveKit room
→ speech-to-text → your grounded release → per-sentence safety pass → text-to-speech
→ the caller hears the first sentence while the next one is still generating

Telephony: your existing numbers connect over an elastic SIP trunk into a LiveKit media room — one room per call, with a persistent cloud agent worker that answers whether or not anyone on your team has a laptop open.

The brain: the agent speaks to your release through an OpenAI-compatible endpoint that fronts Divinci's grounded chat pipeline — RAG retrieval across your vectors, your fine-tuned model, model fallback chains, and signed transcripts, identical to web chat.

The voice: telephony-tuned speech recognition and natural text-to-speech, with barge-in — callers can interrupt mid-answer, like a real conversation.

Built for real conversations

GROUNDED

Your knowledge, not the model's guess

Every answer is retrieved from your release's knowledge base and generated by your configured model — including fine-tunes. If your web assistant recommends against olive oil, so does your phone line. Same brain, different wire.

STREAMING

Sentence-by-sentence speech

The model's answer is streamed, split into complete sentences, safety-checked per sentence, and spoken as it's produced. The caller hears sentence one while sentence two is still generating — no long dead air before a wall of speech.

SAFE

Guardrails before the words are spoken

Each sentence passes your release's safety-rewrite rules before it reaches text-to-speech. For regulated domains — health, finance — non-compliant phrasing is rewritten or suppressed mid-stream, not after the caller has already heard it.

ROUTED

One number per release

Phone numbers map to releases. Run one line for your flagship assistant and another for a different product or brand — each call is routed to the release that owns the dialed number, with graceful fallback.

REMEMBERS

Multi-turn call memory

Ask about blueberries, then "what about strawberries?" — the agent keeps per-call conversational context with cryptographically signed transcripts, the same tamper-evident thread model as anonymous web chat.

METERED

Wallet-bounded, rate-limited

Every turn settles through the release wallet with escrowed cost estimates, per-call turn caps, and a release allowlist — a phone line can never run up an unbounded bill.

Engineered for speed

We measured every phase of the call path and cut the ones that didn't earn their latency — reasoning-phase elimination on flash models, per-sentence streaming, and concurrent cache lookups. The numbers below are from our staging test line.

~4.5sfirst audio on a novel question

<0.5srepeat questions (semantic cache)

~35%generation latency cut by disabling the model's thinking phase for voice

100%of sentences safety-checked before speech

Measured on a RAG-grounded, fine-tuned release answering novel questions end-to-end (retrieval + generation + safety pass). Simple FAQ-style deployments are faster still.

Hear it in action

A lightly-trimmed exchange from our staging test line, running a nutrition-expert release. Note the answer is the expert's position — grounded in the release's knowledge base — not a generic model's.

CALLERIs olive oil healthy?
AGENTNo, olive oil is not a health food; it's very calorically dense, contributes to weight gain, and can raise your LDL cholesterol. While a tiny amount might be okay for someone very thin and active, most people should avoid it.
CALLERI take blood pressure medication. Is it safe for me to start eating lots of leafy greens?
AGENTLeafy greens are wonderfully beneficial for blood pressure, and I highly recommend them, but please discuss your dietary changes with your doctor so they can monitor your medication needs.

Bring your own number

Already have a business number your customers know? Adopt it into a release without porting — the number keeps working, and now it answers. New numbers can be provisioned per release, with SMS and MMS support on the same number: callers can text your assistant a question or even a photo, and get a grounded reply with cited sources.

Put your assistant on the line

Voice Agents are in early access. If you want your knowledge base answering a phone number — grounded, guardrailed, and fast — we'd love to set up a pilot with you.

Book a demo call

AI Voice Agents — Your Grounded Assistant, On a Real Phone Number