Mofid Securitiesdistributed systemsVoIP / SIPapplied AI

Smart Outbound Call Service Automated, Interactive Outbound Voice Platform

Thousands of customer calls, personalised and adaptive, with no agent on the line.

My roleBuilt the controller service; designed the inter-service protocol

CompanyMofid Securities, Iran's largest brokerage

DomainCustomer communications

ScaleHundreds of outbound calls per hour, on demand

01Overview

Mofid regularly needs to deliver important reminders and account updates to large numbers of customers, work that previously required call-centre agents to dial thousands of people by hand. The Smart Outbound Call Service automates this entirely: it places an outbound call to a target customer, runs a defined interactive call flow, plays personalised audio, and adapts to what the customer says, delivering the information without an agent on the line. It removed a large, repetitive manual workload and saved tens of hours of agent call time.

02Architecture

The system is split into two services with a deliberately clean separation of concerns. A stateless Go service registers with the SIP server and handles the live media, sending and receiving audio over RTP. A stateful FastAPI controller is the brain: it holds the call flows, the customer list, and each call's progress, and drives the Go service over a WebSocket action protocol (initiate call, play audio, listen, end call).

Because all state lives in the controller, the media service stays simple and disposable. The controller also owns the intelligence: TTS turns each message into speech, ASR transcribes the customer's reply, and an LLM interprets that reply to choose the next branch in the flow.

Figure 1. Stateful FastAPI controller orchestrates a stateless Go media service over a WebSocket protocol; AI services drive the conversation.

03A single call, step by step

Figure 2. The interactive loop: the controller speaks, listens, interprets the reply with ASR + LLM, and branches the flow accordingly.

04My contributions

Built the controller service end-to-end (FastAPI): the stateful brain holding call flows, customer lists, and per-call progress, and orchestrating every call.
Designed the WebSocket protocol between the controller and the Go media service: the action guideline (initiate call, play audio, listen, end call) that let the Go side stay completely stateless.
Integrated the AI layer into the flow: TTS to voice each message, ASR to transcribe customer responses, and an LLM to interpret those responses and decide the next step.
Contributed to debugging the Go SIP/RTP media service.

Design highlight. Splitting the system into a stateless real-time media layer (Go) and a stateful orchestration layer (FastAPI), decoupled by a small WebSocket action protocol, kept the hard real-time code simple while all conversational logic and state stayed in one testable place.

05Scale & technology

Used on demand for campaigns, the service handles hundreds of outbound calls per hour.

Go (SIP + RTP)Python · FastAPIWebSocketsSIPRTPTTSASRLLMs

06Impact

Replaced manual mass-calling by agents, saving tens of hours of call time per campaign.
Delivered adaptive, LLM-driven call flows rather than rigid menu-based IVR.
Scales to hundreds of calls per hour on demand, with no agent in the loop.

Architecture described at a conceptual level; internal identifiers and credentials are intentionally omitted.