A contact centre that answers in your customer's dialect
Voice, WhatsApp and autonomous agents on one routing engine. The flow engine opens the media and inference planes only when a node needs them, and the whole stack — recognition included — runs inside your own data boundary.
audio stream plane · inference core · agent surfaces — drag to rotate
Your audio does not have to leave the country
Databases, media servers, object storage, the event bus and the inference tier all deploy on infrastructure you control. Channel providers are the only mandatory egress; recognition, generation and synthesis each have an in-boundary counterpart selected per flow.
UAE · Dubai
Primary tenancy. Media, recognition and recordings stay on in-country hosts.
Qatar · Doha
Second hosting region for tenants under Qatari data rules.
Zero audio egress
Select the self-hosted recognition tier and no call audio crosses your perimeter.
Per-tenant keys
Credentials and channel tokens encrypted per tenant, rotated outside the image.
Three parts, one contract
Ittisal Studio
Seven archetypes across seven channels. The archetype decides which nodes exist, the channel removes the rest, and an AI co-pilot builds the first draft of the graph from a written brief.
Voice AI engine
Audio forks off the live call leg and streams into recognition while the caller is still speaking. Managed or self-hosted recognition, generation and synthesis, chosen per flow.
Omnichannel routing
One capacity budget per agent, spent by channel: a call takes all of it, a WhatsApp thread a quarter, an email ticket a tenth. Skills and language are matched before assignment.
A production graph, mid-call
pulses follow execution orderFrom ringing to resolution
trunk · recognition · orchestration · desktop — drag to rotate
Caller arrives
The trunk hands the call to the media server, which asks the flow engine what to do with this number.
Speech to text
Reaching a voice-agent node opens an audio fork. Each leg streams separately, tagged by speaker.
LLM orchestration
Intent, tenant knowledge and guardrails decide the turn. Tool calls hit your CRM through the graph, not a side channel.
Agent desktop
On escalation the agent receives the live transcript, sentiment and a disposition draft — no repeated questions.
Tenant context is forced, not trusted
Every request resolves its tenant once and then runs under it. A hint that disagrees with the session is refused and logged; only platform system roles may cross a boundary, and the target is verified before the request proceeds.
Request the security packEncrypted audio streams
Call legs and recognition forks are carried over TLS; recordings are encrypted at rest in your own object store.
Local GCC hosting
Deploy on in-country infrastructure, including the inference tier. No component requires an external region to function.
Zero data egress
With the self-hosted recognition and generation tiers selected, the only outbound traffic is to the channel providers your customers already use.
Quote-led licensing, metered usage
Enterprise licence
Priced per deployment on residency model, channel mix and seat count. Issued and enforced through the licence server.
Usage on top
Per-minute and per-message rates that step down with volume. Metering derives from the flow graph, so a line on the invoice maps to the nodes that ran.
Deflected minutes cost a fraction of agent-served ones, and the self-hosted tier turns a variable provider cost into fixed infrastructure at scale.
See it answer a live call
A scoped pilot on one queue, with your dialect, your CRM and a success metric agreed in advance.