Technical overview

Frontier’s technical architecture is engineered for real-time performance and scalability, leveraging Cloudflare’s global network to deliver instant sales coaching. At its core, Frontier processes live call audio, transcribes it, infers coaching signals, and delivers these insights to the sales representative’s desktop in milliseconds.

Overall Architecture

Frontier’s system begins on the sales representative’s computer with a desktop application, then extends to a globally distributed Cloudflare edge network that performs the real-time AI processing, before returning signals to the representative. It integrates with external services for meeting capture, speech-to-text, and sophisticated AI models, while managing application data in a secure, scalable database layer. This design prioritizes speed and reliability, crucial for in-the-moment coaching.

⤢ Expand

flowchart TB
rep["Sales rep — the user"]
admin["Sales leader / admin — the buyer"]
frontier["Frontier<br/>real-time AI sales-call coaching"]
recall["Recall Desktop SDK<br/>on-device audio capture"]
deepgram["Deepgram<br/>speech-to-text"]
llm["LLM providers<br/>Anthropic / Google / OpenAI / Together"]
cf["Cloudflare<br/>edge compute + data"]
supa["Supabase<br/>config, jobs, realtime subscriptions"]
rep -->|"sees coaching signals in the HUD"| frontier
admin -->|"sets up scripts, knowledge, teams"| frontier
frontier -->|"captures call audio on the rep's device"| recall
frontier -->|"transcribes the two audio feeds"| deepgram
frontier -->|"generates answers and signals"| llm
frontier -->|"runs on / stores data"| cf
frontier -->|"reads/writes app data"| supa

The Desktop Client

The primary interaction point for a sales representative is the Desktop client, an application built with Electron for macOS. This client hosts the Heads-Up Display (HUD), a real-time overlay that shows coaching cues and information during a call.

For capturing call audio, the Desktop client currently integrates with the Recall Desktop SDK. This SDK runs directly on the representative’s machine, allowing Frontier to receive call audio and participant information. Frontier is actively migrating to a direct audio capture method, where the Desktop client will record the representative’s microphone and the system’s output audio separately. This move will enable complete isolation of the two audio feeds (representative and customer), which is essential for more advanced coaching capabilities.

Real-time Coaching Services (Cloudflare Edge)

The core of Frontier’s real-time intelligence runs on an orchestrated set of Cloudflare Workers. These are globally distributed serverless functions that bring computation close to the user, minimizing latency. Within this architecture, Cloudflare Durable Objects are key. Each active sales call is managed by its own Durable Object instance, providing a unit of statefulness and concurrency that ensures isolated and consistent processing for every call.

When a call is in progress:

1. The audio captured by the Desktop client is streamed to these Cloudflare Workers.
2. From there, it is immediately sent to Deepgram, a specialized speech-to-text provider, which converts the audio into a word-level transcript.
3. This transcript is then fed back into the Durable Object managing the call, where real-time AI models analyze each utterance (spoken segment).
4. These models detect various sales-relevant signals, such as questions asked, compliance risks, or opportunities for script adherence. This inference happens rapidly, often leveraging an orchestrated set of companion Cloudflare Workers for specific tasks like script completion, FAQ detection, and question detection.
5. The detected signals are then delivered back to the Desktop client’s HUD over a secure WebSocket connection, appearing as instant coaching for the sales representative.

GAP — founder supplies

Specific latency, throughput, and concurrency numbers for the real-time processing path are not available in the current documentation. These figures are supplied by the founder and are critical for understanding the system’s performance at scale.

Knowledge and Data Management

Frontier uses a combination of data stores tailored for different purposes:

• Supabase (Postgres): This managed PostgreSQL database is the system of record for application and account data. It stores information about organizations, users, call metadata (but not raw transcripts), sales scripts, FAQs, and knowledge documents. Supabase is used for configuration, background jobs, and its real-time subscription feature, with infrequent calls from the latency-sensitive real-time path.
• Cloudflare D1: This is a SQLite database running on Cloudflare’s edge network. It stores the live call content, including detailed word-level transcripts (containing raw PII) and extracted questions. D1’s proximity to the real-time processing ensures fast access to this critical call data.
• Cloudflare AI Search & R2: Frontier’s knowledge retrieval system is designed for flexibility and is currently in active evaluation with a multi-backend approach. Cloudflare AI Search (an AutoRAG service) is a primary backend for retrieving relevant information from an organization’s knowledge base, backed by Cloudflare R2 for storing the knowledge source files.
• Supermemory: This serves as a deliberate stop-gap knowledge backend, providing an additional layer for information retrieval while the platform evolves.
• Cloudflare Vectorize: While present as a vector index binding, its primary role in live knowledge retrieval has been superseded by Cloudflare AI Search and other systems.
• LLM Providers: Frontier integrates with various external Large Language Model providers (including Anthropic, Google, OpenAI, and TogetherAI) via the Vercel AI SDK. These models are used for more complex inference tasks that generate coaching signals and responses.

The knowledge base layer is an ongoing area of investment. While Cloudflare AI Search and Supermemory are in use today, Frontier plans to experiment with advanced Graph RAG (Retrieval-Augmented Generation) services for even greater accuracy and completeness in knowledge retrieval.

Asynchronous Workflows

While the real-time path is optimized for speed, many critical background tasks are handled by Inngest. Hosted within Frontier’s Dashboard application, Inngest provides a durable asynchronous layer for:

• Processing webhooks from external services like Recall.ai (for meeting events).
• Managing the lifecycle of calls and virtual meeting bots.
• Ingesting and processing new knowledge documents.
• Running scheduled jobs, such as system cleanups and periodic data synchronization.

By offloading these tasks to a separate, durable system, Frontier ensures that the real-time coaching engine remains lean and focused on its core mission of instant feedback during calls.