VoiceCosmos
ARIA answers every booking enquiry, day and night
For: Hotel owners and general managers first, then independent restaurants, spas and salons, and events and catering businesses that lose bookings when nobody answers.
VoiceCosmos is a voice-first operating system for hospitality. Its assistant, ARIA, answers calls, web and voice enquiries for a hotel at any hour, in the customer's language, using what she has learned from the hotel's own website, and turns them into bookings.
It is built for hotels first, with restaurants, beauty and wellness, and events and catering as the other verticals shown on the site. Staff can also run the hotel by speaking: say what you need, and the matching screen opens.
The problems it targets are concrete: after-hours enquiries that ring out and book elsewhere, group proposals that take days to send, and the lack of any record of who asked for what and what was done. ARIA records every step: who asked, what she did, and the system's answer.
VoiceCosmos is sold through a 90-day pilot measured on the hotel's own numbers. ARIA introduces herself as an AI assistant at the start of every conversation, as the EU AI Act requires, and the pipeline keeps a human in the loop: voice in, agent executes, human checkpoint, final output.
€15–€7
per room per month, by room count
€150
minimum per month, after the pilot
90 days
pilot measured on your own numbers
2 of 3
hospitality automations built and in testing
What it does
24/7 guest requests: ARIA takes call, web or voice enquiries, checks rooms, books and confirms by email
Group reservations: group or corporate enquiries become a priced proposal from the hotel's real room rates
Setup by conversation: ARIA crawls the hotel's website and configures flows from a short talk, with no forms
Voice-driven operations: staff speak a request and the right screen opens
A full record of every action: who asked, what ARIA did, and the system's response
Human approval in Pulse for actions that matter during the pilot
Six Day Zero agents per tenant: ARIA, SAGE, NOVA, LUMI, VEGA and ORION
PII redaction on voice transcripts before they reach the LLM or are stored
How we architected it
- Next.js 15
- React
- TypeScript
- Supabase
- Vercel
- VAPI
- ElevenLabs
- Deepgram
- Anthropic Claude
- OpenAI
- xAI Grok (realtime voice)
- RuVector
- Pinecone
- LangChain
- Twilio
- Stripe
- Sentry
- Upstash Redis
- Firecrawl
- 01
One voice runtime (the kernel) runs the same listen, reason, act and learn loop on every surface; industry personas change the voice register, not the kernel
- 02
Pipeline: voice in, agent executes, human-in-the-loop checkpoint, final output; each automation is defined as trigger, acting agent and result
- 03
Next.js application backed by Supabase; the central XENOS CRM (number7evencrm) serves voicecosmos.com, and the hospitality deployment runs as a separate fork (VC-Hospitality) on hospitality.voicecosmos.ai
- 04
Voice stack combines realtime voice vendors with failover, VAPI for phone lines, and ElevenLabs for speech output, with Claude and OpenAI models behind the reasoning
- 05
Each tenant has its own RuVector RAG namespace and database tenant, resolved from the signed-in session
- 06
PMS, POS and CRM adapters authenticate with OAuth2 or signed API keys; outbound calls pass through a vetted MCP boundary
- 07
Privileged actions are logged with timestamp, actor and reason through the Viwago audit trail; voice biometrics come from VoiceIDVault, storing voiceprints as vectors rather than raw audio
- 08
Billing runs on Stripe with a durable event ledger and an activation queue with retries
Want something built like this?
Every build runs the same five gates: spec, architect, build, review, ship. It starts with a 48-hour audit.