AI Voice Calling at Scale

Overview

A multi-tenant AI voice platform running 50,000+ calls per week, built for both inbound and outbound use cases. The system serves 20+ distinct client tenants on shared infrastructure while keeping their prompts, data, and phone number pools fully isolated.

The problem

Several clients needed AI voice agents that worked reliably at production volume — not the demo-ready version that crumbles at 1,000 concurrent calls. Each client had different personas, different knowledge bases, different CRMs to write back to, and different compliance requirements. Building a one-off system per client was not economically viable; building a shared platform with full tenant isolation was.

What we built

• Designed a multi-tenant architecture with per-tenant prompt configuration, knowledge bases, phone number pools, and CRM integrations — but a single shared runtime to amortize operational cost.
• Built on VAPI for the voice agent runtime, Twilio for telephony, Supabase as the data plane (tenant config, call logs, transcripts), and n8n for orchestration + CRM write-backs.
• Engineered for scale from day one: queue management for outbound dialer campaigns, concurrency caps per tenant, intelligent retry + failover, and real-time observability on every call.
• Added compliance guardrails directly into the outbound path: TCPA-aware time windows, DNC scrubbing, consent capture, and jurisdiction-aware recording disclosure.
• Instrumented every call with transcript + metadata for daily quality review, feeding continuous prompt and knowledge-base improvements.

Outcomes

• Platform now handles 50,000+ calls per week across inbound and outbound workloads.
• 20+ clients live on the shared infrastructure, with new tenants able to stand up in about a week.
• Per-call cost is dominated by telephony + model inference — the operational overhead is amortized across the full tenant base.
• Quality on a per-tenant basis improves continuously via the daily review and prompt-update loop, rather than degrading with scale.