Today's sales-team action brief — synthesized from live competitor pages by a Cloudflare Workflow.
Pulling latest run…
Synthesized from live competitor pages via Cloudflare Workflow. Each section links to the source page it was scraped from.
Click Run workflow to fire the Cloudflare Workflow now. It fetches three competitor pages, normalizes signals via Workers AI, and synthesizes the brief.
What's running under the hood and what role each service plays in the pipeline.
The state machine that runs the brief every morning. Each step persists state and retries independently — a failed competitor fetch doesn't kill the run.
Fires the workflow every morning at 10:00 UTC so the brief is ready before the sales team starts.
Real headless Chrome instances fetch the three competitor pages, including SPA-rendered content.
Runs both the structured signal extraction (per competitor) and the brief synthesis when Workers AI is the selected provider.
Every model call routes through one gateway — Workers AI and Claude both. Cloudflare-managed Anthropic key means no separate provider account.
Raw competitor HTML + normalizer raw output + final brief JSON, all stored forever with no egress charges, ever.
Every run, every signal extracted, every model call with cost and latency — a queryable record of everything the agent did.
Holds the competitor URL list, the prompt versions, and other config that should be editable without redeploying the Worker.
One Worker handles cron, the API endpoints, the dashboard assets, and the orchestrator — all from one ~600-line app.
Real-time stack traces and console output. Every wrangler tail session reads from here. Built-in, no setup.
The avalon.fpdigital.ai hostname is gated by Cloudflare Access. Only allowlisted emails reach the dashboard, enforced at the edge before any compute runs.
Define buyer, sales-team action, source boundaries, success metric.
Public pages only, robots respected, no auth, no PII.
Browser Rendering pulls HTML; R2 archives the snapshot.
Workers AI Llama 70B extracts signals into a common schema.
AI Gateway routes synthesis (Workers AI 70B or Claude Sonnet).
R2 brief, D1 audit, Access gate, Workers Logs.
The competitor pages this workflow fetches every morning. Pulled from /api/methodology — exactly what the Worker hits.
How the normalizer classifies every signal it extracts.
How the synthesizer rates each competitor's section in the morning brief.
The exact prompts running in production right now. Pulled from /api/methodology — source of truth.
A daily competitive intelligence agent for the Empire Avalon sales team. Every morning at 6 AM ET, it visits three real competitor builder pages — Losani, Mattamy, Branthaven — and produces a sales-ready brief that lands before the team's first coffee. Same-morning intel. Different than yesterday's static report.
One Cloudflare Worker. A daily cron trigger spins up a Cloudflare Workflow — a durable, retryable state machine. Each step persists independently: if Mattamy times out, Losani's work isn't lost. The whole pipeline runs in 30–60 seconds. Every model call routes through one AI Gateway for unified logging, caching, and cost tracking.
Built by ZeroArc (Fp Digital's Cloudflare MSSP practice) for Empire Homes as the marquee use case of the proposed AI lab. Demonstrates the entire Cloudflare developer stack — Workers, Workflows, Browser Rendering, Workers AI, AI Gateway, R2, D1, KV, Cron, Logs, and Zero Trust — in one ~600-line app.
Llama 3.3 70B for both normalization and synthesis. Averaged from — completed runs in D1.
Synthesizer routed to Claude Sonnet via AI Gateway Unified Billing. Averaged from — completed runs.
Projection: 100× usage growth = ~25× monthly bill. Linear scaling. R2 egress is $0 forever, no matter how many briefs Empire downloads. Directional — based on published Cloudflare unit pricing.
Sandbox-deployed under a separate Worker, D1, R2, and AI Gateway. Application Security gated. Full audit trail. Read the build log →