Voice AI engineering prototype

Every conversation
should be accountable.

Vocalis explores the systems behind a business phone agent: tenant-scoped calls, grounded knowledge, confirmation-gated booking, escalation, and inspectable events.

Public project preview · Live voice demo hosted on Render.

CONVERSATION TIMELINE
01 Speech detected
02 Tenant knowledge retrieved
03 Response and tool events recorded
Illustrative flow · no live call or latency shown

Built around operational reality.

The current repository is a working pilot with explicit boundaries. These pieces are implemented and covered by deterministic tests.

◎

Browser conversation

Microphone and text input over WebSockets, with cancellation of in-flight turns and browser playback.

↗

Inbound phone adapter

Signed Twilio speech webhooks, retry handling, and a configurable human transfer number.

▦

Tenant-scoped records

Persistent calls, events, documents, appointments, and leads with scoped API access.

⌕

Knowledge retrieval

Tenant-isolated text retrieval with source metadata and conservative fallback when facts are unavailable.

✓

Safer booking

Voice bookings require a caller confirmation and a successful calendar operation before confirmation.

◇

Offline evaluation

Deterministic isolation, interruption, webhook, and booking tests plus seeded conversation fixtures.

Traceable from speech to response.

A small service architecture keeps provider calls and tenant data behind explicit boundaries.

01 · INPUTBrowser or phone
02 · VOICETranscribe speech
03 · CONTEXTRetrieve knowledge
04 · DECIDEGenerate or escalate
05 · RECORDPersist call events

FastAPI + WebSockets · Groq speech and language · Edge TTS · Neon Postgres · optional Google Calendar and HubSpot adapters

Try the live Vocalis voice demo

The free backend may need a moment to wake. The demo asks for consent before a conversation begins.

Checking backend