Skip to content
Frontier Docs

System overview

Frontier is a real-time AI sales-call coaching system that provides live guidance to sales representatives during calls. This overview describes the high-level architecture, outlining the core applications, data stores, workflow management, and integrations that enable instant feedback and coaching on the sales floor.

flowchart TB
subgraph client["Client · rep's Mac"]
  desktop["Desktop app<br/>Launcher + HUD"]
end
subgraph edge["Cloudflare edge"]
  cs["Call Server<br/>workers + per-call Durable Objects"]
end
subgraph app["Application plane"]
  dash["Dashboard<br/>Next.js · Vercel"]
  inngest["Inngest<br/>durable jobs + cron"]
end
subgraph data["Data & storage"]
  supa[("Supabase<br/>config · accounts")]
  d1[("D1<br/>transcripts")]
  ais[("AI Search + R2<br/>knowledge")]
  kv[("KV · Vectorize")]
end
subgraph ext["External"]
  recall["Recall<br/>Desktop SDK"]
  dg["Deepgram<br/>speech-to-text"]
  llm["LLM providers<br/>via AI Gateway"]
  clerk["Clerk<br/>auth"]
end
desktop -->|"WebSocket"| cs
desktop -->|"capture"| recall
desktop -.->|"audio (direct)"| dg
desktop -->|"provision"| dash
desktop -->|"sign-in"| clerk
cs --> d1
cs --> ais
cs --> kv
cs -->|"LLM"| llm
cs -->|"transcripts"| dg
cs -.->|"config"| supa
dash --> supa
dash --> inngest
inngest --> ais
classDef cl fill:#e3f2fd,stroke:#1565c0,color:#000
classDef ed fill:#fff3e0,stroke:#e65100,color:#000
classDef ap fill:#ede7f6,stroke:#5e35b1,color:#000
classDef st fill:#e8f5e9,stroke:#2e7d32,color:#000
classDef ex fill:#fff8e1,stroke:#f9a825,color:#000
class desktop cl
class cs ed
class dash,inngest ap
class supa,d1,ais,kv st
class recall,dg,llm,clerk ex

Application Layer

Section titled “Application Layer”

The Frontier application layer consists of the Desktop app and its embedded HUD, the Dashboard for administrative functions, and the Call Server that manages real-time call processing. The Desktop app, an Electron client, runs on the sales representative’s machine, providing the primary user interface and capturing call audio using the Recall Desktop SDK. The HUD (Heads-Up Display) is a Vite/React web app embedded within the Desktop app, displaying live coaching. The Dashboard is a Next.js application that handles provisioning and asynchronous backend tasks. The Call Server is an orchestrated set of Cloudflare Workers and per-call Cloudflare Durable Objects, acting as the edge backend for all live call logic. Learn more about these components on the dedicated Desktop app and Call server pages.

Data & Storage

Section titled “Data & Storage”

Frontier utilizes a diverse set of data stores tailored for performance and scalability. Supabase (Postgres) serves as the primary relational database for application and account data, enforcing tenant isolation through Row Level Security (RLS). Cloudflare D1 (SQLite) stores live-call metadata, including raw word-level call transcripts and extracted questions. For knowledge retrieval, the system leverages Cloudflare AI Search backed by Cloudflare R2 object storage, with Supermemory also in use as a stop-gap solution. Cloudflare Vectorize is present for embeddings, and Cloudflare KV (Key-Value) provides caching for elements like scripts. Explore these data mechanisms in detail on the Data & storage page.

Workflows & Webhooks

Section titled “Workflows & Webhooks”

Asynchronous processing and scheduled tasks are managed by Inngest, a durable workflow engine integrated into the Dashboard. Inngest handles long-running operations such as processing call lifecycle events, knowledge ingestion, and periodic system cleanups via cron-scheduled functions. Historically, the system received meeting events via webhooks from Recall.ai, which were then forwarded to Inngest for processing; this original cloud-bot and webhook path is now deprecated in favor of direct desktop capture. Detailed information on these processes can be found on the Workflows & webhooks page.

Integrations

Section titled “Integrations”

Frontier integrates with several critical third-party services to deliver its real-time coaching capabilities. Recall.ai is used for meeting capture (via its Desktop SDK) and calendar synchronization. Deepgram provides speech-to-text transcription, with Frontier actively migrating to a direct audio capture path for enhanced speaker separation. Large Language Models (LLMs) from providers such as Anthropic, Google, OpenAI, and TogetherAI are orchestrated via the Vercel AI SDK and accessed through Cloudflare Workers AI. Clerk handles authentication, while Doppler manages secrets across environments. Observability is provided by Axiom for logs and traces, Sentry for error reporting, and PostHog for product telemetry. A comprehensive list of these external services and their roles is available on the Integrations page.