Nawa

The problem

A call-center platform serving multiple tenants needs every conversation, tool call, and invoice scoped to the right tenant, with no leakage between them.

Callers arrive two different ways, PSTN through Twilio and browser-based WebRTC, and both need to reach the same AI conversation pipeline.

An AI agent that can take real actions needs those actions gated by policy and logged, since a bad tool call has real consequences for a tenant's operations.

Usage has to be metered accurately enough to drive billing: pricing logic, top-ups, billing cycles, and invoice lifecycle, not just logged for observability.

Role & responsibilities

API architecture across a layered Domain / Application / Infrastructure / API structure in ASP.NET Core.

JWT and API-key security with tenant-aware authorization and middleware policy enforcement.

Real-time interaction flows: chat, speech-to-text, text-to-speech, WebRTC SDP exchange, and WebSocket-based telephony bridging with OpenAI and Twilio.

Policy-driven tool orchestration, webhook outbox/retry, and Hangfire background jobs for auditable external actions.

Billing and monetization: usage metering, pricing logic, top-up/payment flows, billing cycles, invoice lifecycle.

Two call paths, one pipeline

PSTN callers via Twilio and browser callers via WebRTC needed to reach the same AI conversation logic without two divergent implementations. WebRTC SDP exchange and WebSocket-based telephony bridging act as two entry points feeding the same speech-to-text / AI / text-to-speech pipeline, so one conversation implementation serves both caller types.

Letting an AI agent take real actions safely

Tool calls triggered by the AI agent have real side effects on tenant systems, so they can't be executed blindly. Policy-driven tool orchestration gates which tools an agent can invoke, with webhook outbox/retry and Hangfire jobs making external actions auditable and retry-safe.

Metering usage precisely enough to bill

Usage metering and pricing logic are first-class parts of the domain, feeding top-up/payment flows, billing cycles, and invoice lifecycle controls: a billing layer driven by real usage data rather than approximated after the fact.

Outcome

A production-grade AI operations layer that connects real-time voice and chat to billable, tenant-aware infrastructure, handling both traditional telephony and browser-based calling behind one conversation pipeline.

What I learned

Bridging two caller protocols into one pipeline is more work upfront than treating them separately, but it means the AI logic only has to be built once.

Billing logic belongs in the domain model from day one on a metered platform: retrofitting usage metering onto an existing conversation flow is much harder than designing for it.