Session #18271 Live transcript
Kasipan VoiceAgent · v4.2
00:24 / 02:11
— 287ms latency · 22kHz · Opus
Transcript EN-US · auto
Cost / minute
$0.018
CSAT (last 30d)
4.6 / 5
Containment
82%
Avg. handle time
−38%
[ Capabilities / 008 ]

Engineered for
the phone call, not the demo.

→ 001 / TURN-TAKING
Predictive endpointing.
Our endpoint model fires ~120ms before silence, then negotiates barge-in with the user's prosody. Conversations don't stutter.
→ 002 / VOICE
Streaming TTS, custom-cloned.
Bring 90 seconds of audio. We clone, polish, and stream at 22kHz with phoneme-level emotion control. Brand-safe by default.
→ 003 / TELEPHONY
Drop-in to Twilio, Vonage, your PBX.
SIP-native. Works with the numbers you have. Failover routes to a human in < 2s if confidence drops.
→ 004 / TOOLS
Calls your stack, not the other way around.
Native function calling against CRM, scheduler, billing, ticketing. Tools are typed; eval harness rejects builds that hallucinate arguments.
→ 005 / OBSERVABILITY
Every turn, traced.
OpenTelemetry spans for ASR, LLM, TTS and tool calls. Replay any conversation as a flame graph. Diff prompts across deploys.
→ 006 / SAFETY
PII redaction at the edge.
Card numbers, SSNs and PHI scrubbed before transcripts leave your VPC. SOC 2, HIPAA-ready.
→ 007 / EVAL
10k synthetic callers, nightly.
Adversarial test suite runs against every PR. We catch the regression before your customer does.
→ 008 / DEPLOY
Two-week production cutover.
Discovery → prototype → shadow mode → live. We run all four in parallel with your existing line.
[ Pipeline / Realtime ]
Telephony SIP / Twilio VAD endpointer ASR streaming Reasoner tool-calling LLM Tools CRM · Sched · Pay TTS streaming clone Caller 22kHz Opus ~12ms ~70ms streaming ~95ms first token ~110ms TTFB Total: 287ms p50 · 412ms p95
[ Specifications ]

Numbers we'll
put in the contract.

→ Latency
287 / 412 ms
First-byte audio response, p50 / p95. Measured end-to-end across the pipeline.
SLA · 350ms p50
→ Containment
82%
Calls fully resolved without human handoff, averaged across active deployments.
Range 71–94%
→ Languages
34 / 60+ accents
Native speech models for English, Spanish, French, Hindi, Tagalog, Arabic, plus 28 more.
Custom on request
→ Cost
$0.018 / min
All-in compute, including ASR, LLM, TTS and tool calls. Volume discounts at 100k min/mo.
vs. $0.95 human
→ Compliance
SOC 2 · HIPAA · PCI-DSS
Edge PII redaction, BYOK encryption, region-pinned data residency. Audited annually.
Evidence on request
→ Deploy time
14 days
Discovery to live-traffic shadow mode. Most teams cut over fully within 21 days.
Median across 31 builds
[ Developer experience ]

A handful of
lines, not a vendor lock-in.

Every Voice AI deployment ships with the spec checked into your repo. Prompts, tools, evals, voices — all versioned, all yours. Run the agent locally, replay any production call, and roll back like any other service.

→ 01Streaming WebSocket and SIP transports
→ 02Typed tool definitions, JSON-schema validated
→ 03Eval harness checked into CI
→ 04Single-tenant deploys to your VPC
agent.spec.ts
main · clean
// Voice agent spec — versioned, replayable, yours.
import { defineAgent, tool } from "@kasipan/voice";

export default defineAgent({
  id: "front-desk-v4",
  voice: "clone:lena/22k",
  latency: { p50: 280, p95: 450 },

  prompt: readSpec("./prompts/front-desk.md"),

  tools: [
    tool({
      name: "book_appointment",
      schema: BookingSchema,
      handler: async ({ customerId, slot }) => {
        return await crm.book({ customerId, slot });
      },
    }),
  ],

  guardrails: {
    redactPII: true,
    handoffOn: ["escalation", "low_confidence"],
  },

  evals: "./evals/*.yaml", // 10k synthetic callers, nightly
});
[ In production ]

Where small teams
are deploying it.

→ 001 / DENTAL · 14 LOCATIONS
After-hours booking,
without a call center.
Front-desk agent answers, books, reschedules and handles insurance lookups. New patient bookings up 3.1× since deploy.
Containment88%
→ 002 / HOME SERVICES
Outbound win-back,
at $0.018/min.
Voice agent reactivates lapsed customers with personalized offers pulled from CRM. Replaces a 6-person team's outbound queue.
Reactivation lift+47%
→ 003 / RECRUITING
Phone screens,
at the speed of apply.
Candidates get a call within 4 minutes. Agent screens for fit, books a recruiter slot, and writes a structured eval into the ATS.
Time-to-screen−92%
[ Pilot program / 2 slots ]

14-day pilot.
Production traffic on day eight.

Bring a single workflow — inbound, outbound, or front-desk. We instrument, build, and ship a measurable agent against it. If it doesn't move a number, you don't pay.