Real-Time AI Sales Coaching for Telemarketing Agents

Project: Real-Time Sales Coaching

Tech Stack: Node.js, TypeScript, React, WebSocket, STT (multi-provider), LLM coaching engine, SFTP deployment

Category: Real-Time AI, Voice Intelligence, Sales Technology

Background

This is a real-time AI-powered sales coaching system built for JustDial's telemarketing team. It listens to live phone calls and delivers instant AI-generated coaching insights to agent dashboards — while the call is still happening.

How It Works

PBX/Dialer (Audio) → WebSocket → coaching-core (Node.js)
                                        ↓
                              Dual-channel STT
                              (agent + customer separately)
                                        ↓
                              LLM Coaching Engine
                                        ↓
                              coaching-front (React Dashboard)
  1. Audio ingestion: Raw PCM audio from PBX via WebSocket (two channels: agent mic + customer)
  2. Speech-to-text: Each channel piped to STT separately — preserves speaker identity
  3. Coaching: LLM analyzes conversation in real time and generates coaching hints
  4. Dashboard: Agent sees live suggestions without the customer hearing anything

Challenge 1: Dual-Channel Audio Processing

Problem: Standard telephony setups mix agent and customer audio into a single stream, making speaker diarization complex. The system needed to identify who said what accurately.

Solution: The PBX was configured to deliver two separate WebSocket streams — one per channel. This eliminates diarization ambiguity and allows the LLM to reason clearly about agent vs. customer intent.

Challenge 2: Remote Deployment via SFTP

Problem: Development was done locally but deployment was to a remote production server (192.168.40.170). Manual file copying was error-prone.

Solution: Used VS Code SFTP extension with a .vscode/sftp.json configuration:

json
{
  "name": "SSO",
  "host": "192.168.40.170",
  "protocol": "sftp",
  "port": 22,
  "username": "...",
  "remotePath": "/var/www/app"
}

This enabled one-click sync of code changes to the production server.

Challenge 3: Understanding the Existing Architecture

Problem: A new developer needed to quickly understand a complex dual-service architecture (backend + frontend, real-time audio + LLM) without documentation.

Solution: Claude analyzed the entire codebase and produced a full architectural breakdown covering: audio flow, WebSocket protocols, LLM integration points, and frontend state management — used as the onboarding document.

Key Learnings

  • Separate audio channels per speaker is far more reliable than post-hoc diarization
  • Real-time coaching must have <500ms latency or agents lose trust in the system
  • VS Code SFTP extension is a practical tool for rapid iteration on remote servers

Session Date: February 2026 | Environment: Production PBX integration