Pepper Ad Coach: AI-Powered Ad Monitoring Platform With Zero Runtime LLM Calls

Unleashed Consulting's media buyers were skilled but mathematically overwhelmed: each of 80+ clients ran on multiple ad platforms, each with its own dashboard. Deep-optimizing one account meant not watching the other seventy-nine. The agency wanted every client to receive expert-level attention on every monitoring cycle — and wanted that coverage to scale as the client book grew without scaling headcount. Existing "AI-powered" ad tools called LLMs on every analysis, producing variable costs, opaque outputs, and fragile testing, none of which were acceptable at agency scale.

Fountain City designed Pepper Ad Coach using a "compiled AI" architecture: LLMs were used heavily at build time to author the system's expert judgment into a Decision Matrix of 15 lookup tables with hundreds of unique cells, but the production system makes zero LLM calls at runtime. Every two hours during business hours, Pepper pulls live campaign data, evaluates it against per-client targets and baselines through pure arithmetic lookups, and produces two output types: instant critical alerts (Slack + email) and batched coaching items ranked by urgency (act today / this week / this month). A self-service onboarding portal lets new clients connect ad accounts without an engineering ticket, and each client's OAuth token is authorized independently for fault isolation. Security was achieved by elimination — zero inbound network ports, Cloudflare Tunnel, encrypted tokens, and a private database with no internet access. The entire platform was built in 11 business days using an agent-driven development pipeline with human review at defined quality gates.

Unleashed Consulting now has whole-portfolio coverage across 80+ clients on a two-hour monitoring cycle — without adding headcount. Campaign problems are caught within roughly two hours of starting, shifting the agency from reactive to proactive. Every client receives the same expert-grade first-pass diagnosis regardless of which media buyer is assigned, and every recommendation traces to an explicit lookup rule and real numbers, making advice defensible to clients. The deterministic architecture keeps runtime costs flat regardless of usage volume, enabling predictable per-client pricing for the agency and a viable margin on a white-label product they can stand behind. The platform was delivered in approximately 11 business days, with the primary human time investment being OAuth app review and platform paperwork rather than coding. Spec conformity on the second pass through all requirements was very high, with only minor issues found.