System Architecture

Overview

Frontier AI is a real-time AI-powered sales coaching platform built as a Turborepo monorepo optimized for fast AI processing during live sales calls. The system integrates with major video conferencing platforms and leverages Cloudflare's edge network for real-time AI coaching.

Key Architectural Principles

• Full TypeScript Stack - Type-safe development across frontend and backend
• React Everywhere - Unified React codebase for web and desktop applications
• Cloudflare Edge Network - Distributed infrastructure with Durable Objects for real-time processing
• AI-First Architecture - Integrated AI workflows for real-time coaching and post-call analysis
• Serverless & Scalable - Pay-per-use infrastructure that scales automatically

System Overview

The platform consists of multiple client applications connecting to a unified API backend, with specialized services for real-time AI processing and post-call analysis.

Core Applications & Data Flow

This simplified view shows how the main applications interact through the API backend:

Key Architectural Decisions

Real-Time vs Post-Processing AI

The system uses two distinct AI processing paths:

1. Real-Time Processing (OpenAI Queue)

   • Direct OpenAI API calls during live calls
   • Fast response times for live coaching
   • Rate limiting and connection pooling
   • Used by Durable Objects for immediate AI responses

2. Post-Call Processing (Humanloop)

   • Transcript analysis after calls end
   • Complex multi-step AI workflows
   • Prompt versioning and A/B testing
   • Managed by Inngest for reliable processing

Monorepo Architecture

• apps/dash - Next.js dashboard with API routes
• apps/web - Desktop app with Neutralino.js
• apps/server - Cloudflare Worker API backend
• packages/ - Shared libraries for types, UI components, database client

Technology Stack Summary