SingaporePanamaHong KongChina

GPU fabric / Singapore Singapore

Platform / Infrastructure

Singapore primary · Customer-selected regions

The infrastructure is the product.

Voice agents are judged in milliseconds and audited in years. The reference architecture covers the compute, media path and models that decide both.

Topology GPU inference · SIP edge · media path · speech models, operated as one system. No mandatory resold model or carrier in the critical path.
4
Layers covered: speech, telephony, compute, ops
3
Deployment models
99.9%
Target availability
24×7
Operations and on-call

01 / Compute

The model is only as fast as the box.

We run inference on current-generation accelerators in Singapore. Pools are sized for the call mix: concurrency, utterance length, language and the number of tools a turn can touch.

  • Continuous batching and quantisation profiles per model.
  • KV-cache control, so a long conversation does not get slower.
  • Dedicated weights and capacity for single-tenant pools.
  • Canary releases, a baseline and a rollback for every model change.

No shared training. Customer recordings are not pooled into a model. Dedicated deployments keep their weights separate.

A dedicated pool is not a shared tenant with a bigger number. It has its own capacity, weights and retention rules.

02 / Telephony

Private edge / numbers

The line is where the truth is.

Most voice agents stop at the API boundary. We operate the SIP edge, the media servers and the carrier interconnects. A call that fails at 2am is ours to fix.

01 / PSTN
Carrier handoff

Local numbers, porting and CLI. Terminated in Singapore or the region agreed in the contract.

02 / SBC
SIP edge

TLS and SRTP, rate limits, toll-fraud checks and an edge the deployment operates.

03 / MEDIA
RTP path

Jitter buffer, echo control and codec negotiation. Traceable per call, per leg.

04 / AGENT
Application

ASR, policy, tools and TTS. The conversation is a state machine, not a chain of prompts.

Codecs
G.711, G.722, Opus
Transport
SIP/TLS, SRTP, WebRTC
Numbers
Singapore and customer-selected regions
Porting
Managed directly
Fraud
Toll-fraud monitoring and spend caps
Emergency
Out of scope; calls are routed away

We do not route emergency calls. Emergency services require capabilities and certifications outside our scope. If a workflow touches emergency dispatch, we say no before a contract is signed.

Number portability is operational work. We handle the paperwork, the cutover window and the fallback. The port is not finished when the form is submitted.

03 / Data

Data follows the contract.

A voice call carries more than audio. It carries identity, intent and sometimes a payment. The data path is agreed before the first call.

  • Encryption in transit and at rest.
  • Data residency in Singapore or the region you choose.
  • Recording optional; off by default unless the flow requires it.
  • PII redaction before storage when configured.
  • Retention and deletion defined in the contract.
  • No training on customer data without written instruction.

Read the data commitments

Data path / excerpt Per deployment
Ingress
TLS or SRTP. Recording is disabled unless the flow and the contract enable it.
Processing
ASR and TTS run in memory. Model calls stay inside the deployment region.
Storage
Transcript and summary. PII redaction runs before the write, not after.
Egress
Signed webhooks to your systems. No third-party analytics or ad pixels.
Deletion
Per-call or per-period, with a deletion record. No soft-delete surprises.

The exact path, retention and subprocessors are written into the deployment record.

04 / Operations

Someone is always on

A distributed system fails in the seams.

We run a 24×7 NOC, an on-call rotation and a written incident process. Response targets are contractual, and the post-incident review is shared.

24×7
Operations and on-call
15 minutes
P1 acknowledgement target
99.9%
Target availability
  • Runbooks for every dependency and every carrier.
  • Game days against the failure modes we actually see.
  • Post-incident reviews with the customer in the room.
  • Capacity review before a launch, not after a queue forms.

Targets, not guarantees. Contracted SLAs, credits and escalation paths are agreed per deployment. The numbers above describe the operating target, not a promise on a marketing page.

05 / Sovereign

Air-gap is a deployment, not a slide.

We ship the model weights, the runtime and the runbooks. Updates arrive as signed bundles. Nothing calls home unless the contract allows it.

Same runtime, different perimeter. The agent behaves the same in our cloud and in your building. What changes is who holds the hardware and the keys.

  • Hardware bill of materials and capacity plan.
  • Signed model and runtime bundles.
  • No egress; telemetry stays inside the perimeter.
  • Your telephony, your number inventory, your carriers.
  • On-site or remote support, agreed in advance.
  • Exit plan: weights, transcripts and configuration are yours.

Discuss a sovereign deployment

Start a conversation

Bring us the call.

Tell us what the call has to do. We will tell you what it takes to run it.

Singapore contracting  ·  Confidential by default  ·  A solutions engineer replies within one business day.