Platform/Architecture

Architecture, MVP & Principles

Modular cloud-native design · MVP monitoring-first scope · phase-two roadmap · acceptance criteria.

OpenAPI

Suggested stack

Aligned with production ZiplyHuman deploy

Frontend

· Next.js · React · TypeScript · Tailwind
· Zustand · Recharts · Accessible UI
· WebSocket/SSE-ready live APIs

Backend (this deploy)

· Next.js Route Handlers + workers path
· PostgreSQL transactional + metrics tables
· Redis cache/queue-ready

Target data plane

· Timescale/ClickHouse analytics
· OpenSearch transcripts
· Object storage recordings
· Kafka/Redpanda events · OTel Collector

Infra

· Docker Compose (live) · K8s/Helm ready
· Terraform · GitHub Actions · secrets/KMS
· CDN · WAF · multi-AZ

Full diagrams, ERD and IaC live under /docs and /db/schema.sql.

Architectural services

API gatewayAuthTenantIntegrationWebhook ingestStreaming ingestNormalizationAudioTranscriptTraceMetricsEvaluationRoot-causeAlertIncidentReportingBillingNotificationAudit

MVP integrations

LiveKitPipecatElevenLabsRetell AIVapiTwilioGeneric webhookGeneric OpenTelemetryCustom REST API

MVP exclusions

· Voice AI agent builder

· Prompt-authoring for production agents

· Contact-centre agent desktop

· Complete carrier network

· Full CRM

· Full ticketing platform

· Automatic production changes without authorization

Phase two

SIPRECDeep SIP diagnosticsRTP quality analyticsWebRTC diagnosticsCarrier benchmarkingHuman QA reviewSynthetic monitoringLoad testingAdvanced anomaly detectionEvaluator calibrationCompliance packsCost optimizationPrivate-cloud deploymentPartner and white-label modeMobile incident applicationAdvanced executive reportsCustomer-managed storageCustomer-managed encryption keys

MVP capabilities checklist

✓ Multi-tenant authentication

✓ Organization and workspace administration

✓ Production-call ingestion

✓ Real-time dashboard

✓ Call Explorer

✓ Call-detail replay

✓ Transcript and recording display

✓ ASR, LLM, TTS and end-to-end latency

✓ Call outcome and success rate

✓ Tool-call tracking

✓ Basic voice-quality metrics

✓ LLM-as-judge evaluators

✓ Custom evaluator builder

✓ Weighted scorecards

✓ Alerts

✓ Slack, Teams and email notifications

✓ Deployment annotations

✓ Basic root-cause analysis

✓ API keys

✓ Audit logs

✓ Usage metering

✓ Subscription plans

Acceptance criteria (demo workflow)

1. Admin creates organization and project

2. Admin connects supported Voice AI platform

3. Production events arrive via webhook/connector

4. Live call on real-time dashboard

5. Transcript and operational events displayed

6. Call ends; ASR/LLM/TTS/total latency calculated

7. Configured evaluators run

8. Call receives quality score

9. Call in Call Explorer

10. Synchronized transcript, audio and trace replay

11. Failed tool call in waterfall

12. Root-cause engine with evidence

13. Threshold breach creates alert

14. Notification sent

15. Incident creatable from call

16. Deployment comparison for regression

17. Usage recorded for billing

18. Sensitive actions in audit log

19. Other tenant cannot access call or related data

Product principles

· Monitoring first, testing second

· Independent of Voice AI vendors

· Observe production agents without migration

· Correlate audio, conversation, AI, app and telecom

· Evidence for every AI diagnosis

· LLM evaluator output is probabilistic

· Never silently hide missing data

· Every metric explainable

· Preserve raw events for authorized investigation

· Protect recordings and sensitive information

· Strict tenant isolation

· Modular integrations

· Fast investigation over decorative dashboards

· Avoid vanity metrics

· Connect technical failures to business impact

· Never unauthorized production changes

· No compliance certification claims without audits

· Enterprise SaaS, not a prototype dashboard