A production-grade customer-experience agent your customers can type to, talk to, or send a photo — grounded on your real systems, never on guesswork. Run it on Google’s frontier models under enterprise terms, on open weights you own inside your own cloud project, or fully on-premise where no cloud is permitted.
Until mid-2026, a capable multimodal CX assistant meant sending customer conversations to a model vendor, token by token. Google’s Gemma 4 release ended that trade-off: a 12-billion-parameter model that hears, sees, and reasons — small enough to run on hardware your project already has access to, licensed so that you own it.
Every figure on this row is verified against Google’s published Gemma 4 developer guide and model card (June 2026) — and re-verified at scoping.
Gemma 4 12B ships under an open licence. No per-token fee to a model vendor, no obligation to send a single customer utterance outside your boundary.
Google’s own guidance: small enough for 16 GB of VRAM. In practice, a single NVIDIA L4 serves it — which changes what “private deployment” costs.
The encoder-free architecture ingests speech and images directly — the largest audio-capable model in the Gemma 4 family. A customer can talk to it or send a photo.
Enough working memory for long support conversations, full policy documents, and the retrieved knowledge that grounds every answer.
The sovereignty ladder
The agent design is identical at every rung — the same tool-grounded loop, the same channels, the same insights. What changes is where the model runs, and who holds the keys.
Frontier capability, enterprise terms.
Gemini on Vertex AI, running in your own Google Cloud project against regional endpoints.
Best for: Retail and commerce CX, and any program where capability speed matters more than weight ownership.
Your data never leaves your project.
Gemma 4 12B — open weights you own — served on GPU inside your own cloud project. No third-party model provider in the inference path.
Best for: Regulated data where cloud is permitted — the tier most sovereignty conversations actually land on.
Your data never leaves your building.
The same client-owned weights on your own GKE cluster or bare metal — air-gap capable, with a fully self-hosted voice stack.
Best for: No-cloud mandates, classified environments, and boards that require physical custody of the model.
Pricing follows the engagement, not a rate card: every tier is scoped and quoted against your integration surface, channels, and regulatory posture during the Assessment.
Most vendors answer the residency question with a hand-wave. We answer it with region codes — including the regions where the honest answer is “not there, use the Sovereign tier.” That precision is the point.
GPU quota is confirmed for your specific project during the Assessment — documented availability and granted quota are not the same thing, and we treat them accordingly.
L4 and G4 GPU capacity is documented for this region. The Private tier runs Gemma 4 12B on GKE inside the Kingdom.
Managed tier serves via regional endpoints. A hard in-country inference requirement routes to the Sovereign tier.
An in-country UAE mandate is answered honestly: the Sovereign tier, on your own infrastructure.
Where in-country is preferred but not mandated, the Private tier can run serverless on Cloud Run GPU in six documented regions.
Reference implementation
“Ask Aqua” runs today on aquora.ae — an Emerge-operated UAE commerce property with a catalogue of roughly six thousand products. It is the Managed tier of this exact architecture, in production: an agentic tool loop in which every product, price, and order status the agent shows is grounded in a live system result. It cannot invent a SKU.
The Private and Sovereign tiers swap the brain for Gemma 4 12B. The proven design does not change.
How we engage
Sovereignty programs fail when the architecture is chosen before the regulation is read. So the engagement starts with a fixed-scope Assessment — and every phase after it is gated, named, and quoted before it begins.
A fixed-scope diagnostic. Output: your regulatory posture mapped, a tier recommendation with evidence, an integration inventory, and the target architecture. Built so procurement can sign it quickly.
The agent built and grounded on your real systems — knowledge base, order or case APIs, escalation paths. Channels to match your posture: web, WhatsApp where permitted, email. Named KPIs and phase gates throughout.
A managed monthly cadence: model updates, prompt and tool tuning, evaluation reviews, and an SLA. Your team keeps custody; ours keeps it sharp.
On the Private and Sovereign tiers — you do. Gemma 4 is released under Apache 2.0, so the weights sit in your own registry, and any fine-tune we build on your data is yours outright. There is no licence to renew and no vendor who can switch it off.
The honest answer: a frontier managed model is stronger on open-ended tasks. A CX agent, however, is not open-ended — it is tool-grounded and narrowly scoped, and the architecture does the heavy lifting. That is why the Assessment includes an evaluation gate on your actual use-cases before you commit to a tier. We publish no benchmark claims; we test on your traffic.
We scope and evaluate Arabic quality per use-case during the Assessment rather than making a blanket claim. If the evaluation gate says the quality is not there for your dialect and domain, we tell you — and design the fallback before launch, not after.
WhatsApp messages transit Meta’s Business Platform by design — for most sovereign deployments that is disqualifying, so we default those programs to web, voice, and kiosk channels. If your regulator signs off on WhatsApp for specific journeys, we wire it. We would rather surface that tension in scoping than discover it in an audit.
The Data-Residency Assessment is a fixed-scope Crawl phase: your regulatory posture, the right tier with evidence, and a procurement-ready architecture — before you commit to a build.