SingaporePanamaHong KongChina

GPU fabric / Singapore Singapore

Platform / Voice agents

Singapore · Reference stack

A script can read. An agent has to listen.

Streaming speech, real barge-in, tool calls and a clean handoff to a human. Designed for private models, telephony and compute, not a mandatory chain of resold APIs.

Reference turn 480 ms from end of speech to first audio. 610 ms at p95. Measured against a written budget, not a demo.
1
Languages, plus code-switching
480 ms
Reference turn, p50
99.9%
Target availability
24×7
Operations and on-call

Specimen / inbound

The turn is the unit.

A voice agent is judged one turn at a time: did it hear the caller, did it answer in time, did it move the task forward? Everything else is a dashboard.

  • Partial speech is processed while the caller is still talking.
  • Barge-in stops playback and discards the queued audio.
  • Tool calls have a timeout, a retry rule and a spoken fallback.
  • Every turn is traceable by call ID, without recordings by default.

Specimen, redacted. The transcript is generated by the agent and faces the same retention rules as production calls.

Specimen / redacted EN · verified line
00:00.412 AGENT Aldaria service desk. This call is recorded. How can I help?
00:02.118 CALLER My internet has been down since—
00:02.640 EVENT barge-in detected · TTS stopped in 48 ms
00:03.002 AGENT I'm sorry. Let me check the line.
00:03.410 TOOL lookup_line(verified) · 312 ms
00:04.020 AGENT I can see an outage in your area. The estimated fix is 16:40.
00:07.884 CALLER Can you text me when it's back?
00:08.120 AGENT Yes. I'll send it to the number you're calling from. Anything else?
00:09.400 CALLER No, that's all.
00:09.720 AGENT Thank you. Goodbye.
Transcript generated by the agent, redacted for publication. Human review is available on every call.

03 / The call

Four movements

A good call has a shape.

The agent does not improvise the structure. It opens, verifies, acts and closes, and knows when to stop and hand over.

01

Greet & disclose

Who is calling, what is being recorded, and how to reach a human. The disclosure is part of the first turn, not a recorded menu.

0.0 s · identity and consent
02

Understand & verify

Intent, entities and authentication. The agent asks one question at a time and reads back the critical values before acting.

1.2 s · intent and identity
03

Act

Tool calls into your systems, with a timeout, a retry rule and a spoken fallback. The caller is never left in silence.

2.8 s · the task
04

Close & hand off

Confirm the outcome, set expectations, and pass the transcript, entities and attempted actions to a human if needed.

12.4 s · outcome and context

04 / Capabilities

Engineering, not theatre

What it does, in the call.

The feature list is short on purpose. These are the behaviours that decide whether a caller stays on the line.

Turn-taking
Streaming ASR, adaptive endpointing, barge-in and backchannels. The agent can be interrupted mid-sentence without losing its place, and it knows when a pause is a thought rather than a turn.
Language
the languages agreed for each deployment, with code-switching inside a turn. Custom lexicons cover product names, addresses and the way your customers actually say them.
Tools
REST, webhooks, CRM and CCaaS connectors. Every tool call is bounded by a timeout and a fallback. If a system is slow, the agent says so instead of pretending to think.
Handoff
Warm transfer with transcript, entities and attempted actions. Queue routing, callback or a human on the line. The next person does not ask the caller to start again.
Post-call
Transcript, summary, disposition codes and QA flags. PII redaction before storage when configured. Retention follows the contract, not a default.
Evaluation
Replayable test calls, a regression suite, WER and MOS tracking. A change ships with a baseline and a rollback, not a hunch.

05 / Integration

A tool contract, not a black box.

The agent can call your systems. Every call has a contract: a timeout, a retry rule, an auth method and a fallback. You can read the trace.

  • Telephony: SIP, WebRTC and PSTN; bring your numbers or port them to ours.
  • Application: REST, webhooks and signed callbacks.
  • Contact centre: CRM and CCaaS connectors, or a direct queue handoff.
  • Data: streaming events, end-of-call summaries and batch export.

No rip and replace. The agent sits in front of the systems you already run. We do not require you to move your CRM or your contact centre to use it.

tools/lookup_order.json contract / excerpt
// bounded, auditable, versioned
{
  "tool": "lookup_order",
  "description": "Return order status for a verified caller.",
  "timeout_ms": 800,
  "retry": 1,
  "auth": "mTLS",
  "fallback": "handoff",
  "log": ["request_id", "latency", "outcome"]
}
Trace Every turn
Call ID
One identifier from edge to outcome
Timings
Per-stage, not a single average
Decisions
Tool, policy and escalation path
Redaction
PII removed before storage

06 / Evaluation

Measured before launch

We replay calls, not demos.

Every change is tested against consented recordings and a synthetic regression suite. The scorecard is agreed before the first production call.

WER
Tracked per locale and domain vocabulary
MOS 4.2
Perceived quality target, evaluated by listeners
Containment
Resolved without a human, by intent
610 ms
Reference p95 turn, measured end to end

No training on customer data. Call recordings and transcripts are used to run and improve your deployment only when the contract says so. They are not pooled into a shared model.

Human escalation is a feature, not a failure. The agent escalates on request, on a failed authentication check, on a policy boundary, or when confidence stays below the agreed threshold.

07 / Guardrails

Voice is a public interface.

A voice agent speaks with your brand and touches real people. The guardrails are designed into the call, not appended to a policy page.

Read responsible voice

  • Disclosure at the start of every call, including recording.
  • Consent captured before any voice clone or outbound campaign.
  • Suppression lists and calling windows enforced before dialling.
  • Human escalation on request, or when a check fails.
  • No emergency services, political persuasion or robocalls.
  • Every call logged, reviewable and deletable under the contract.

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.