Deployment architecture

Your data. Your environment. Your control.

Data residency here is product architecture, not a certification slide. Evaluating a voice model should never require creating a new system that owns your customer conversations.

What you control

Deployment
The data plane runs in the environment you choose: on-premises, in your VPC, or in a customer-controlled cloud account. Nothing depends on a shared multi-tenant service.
Storage
Call media references, transcripts, evaluation runs, and results live in your object store and your database. There is no vendor-side copy.
Retention and deletion
You set retention windows and deletion behaviour across transcripts, media references, evaluation results, logs, and backups.
Credentials
Model-provider and platform credentials stay in your secret manager. VaaniEval reads them at runtime and never transmits them outward.
Provider routing
External model calls happen only through routes you configure, per evaluation run, with a record of which inputs were sent where.
Redaction
Optional redaction policies can remove names, numbers, and identifiers before any input leaves your environment for a hosted provider.

Reference topology

  1. Approved call media and metadata are imported from your voice platform into storage you already own.
  2. An evaluation API and worker run in your cluster, reading that storage over internal network paths.
  3. Candidate models are invoked through connectors you configure: self-hosted endpoints, private endpoints, or approved hosted providers.
  4. Results, traces, and comparisons are written back to your database and surfaced in the review workspace.

A typical first deployment is a small Kubernetes namespace or a pair of VMs with PostgreSQL and an object bucket. We size it with your infrastructure team before implementation begins.

Being precise about “nothing leaves your environment”

When every model under evaluation is self-hosted or reachable on a private endpoint, an evaluation can run with no external network egress at all.

If you choose to benchmark a hosted model provider, that provider necessarily receives the inputs you approve. We will not claim otherwise. What VaaniEval gives you is the control layer around that decision: explicit provider routing, optional redaction, per-run approval, and a record that separates platform data stayed in your environment from approved data was sent to a provider you selected.

Capability status

So a security review starts from facts rather than marketing:

Shared operating responsibilities

You control infrastructure, capacity, networking, retention, and access. VaaniEval provides deployment artifacts, upgrade guidance, health checks, and support procedures, and we agree on ownership before implementation begins.

Two things are worth planning for honestly. Self-hosted deployment lengthens procurement because security and infrastructure teams are involved early, and real-time shadow evaluation increases model-provider spend. Budgets, rate limits, redaction policy, and audit logging belong in the deployment design rather than in a later surprise.

Bring your security team to the demo.

The deployment review is more productive with infrastructure and security in the room. See what the evaluation workspace does.

Own your benchmark

Choose voice models using evidence from your own calls.

See how a private benchmark runs inside your environment, on your production scenarios, across the models you are considering.

Schedule a conversation

Book time with VaaniEval

Choose a time that works for you. You can book without leaving this page.

Calendar not loading? Open the booking page.

Want a quick overview first? Watch the product walkthrough in a new tab.

30 minutes. Bring your model-selection question.