marketingAI

Problem & Operational Bottleneck

Marketing workflows crossed AI generation, brand context, approvals, publishing, and channel integrations. Fragmented scripts could not provide one auditable path from request to reviewed output and asynchronous delivery.

Constraints & Requirements

• Human approval remained mandatory before publishing or external replies
• Long-running generation jobs required durable queue and failure states
• Brand context and structured outputs had to remain consistent across tools
• Production traffic had to move safely across web, API, queue, database, and object storage

System Architecture & Ownership

Built a React and TypeScript web application with a Python FastAPI backend and Pydantic AI agents. The current production topology runs web and API services in Cloudflare Containers, dispatches asynchronous work through Cloudflare Queues, and persists data and media in Neon PostgreSQL and R2.

Engineering Decisions & Rationale

• Used Pydantic AI for typed dependencies, structured outputs, and explicit tool contracts.
• Kept approval state and job history durable in PostgreSQL while Cloudflare Queues handles asynchronous dispatch and dead-letter routing.
• Containerized the web and API services and migrated the production runtime to Cloudflare while retaining documented rollback paths.

Reliability, Retries & Safeguards

• Pydantic request, response, and agent-output schemas make invalid states visible at service boundaries.
• Logfire and OpenTelemetry traces connect application requests to agent execution.
• Cloudflare Queues and its dead-letter queue preserve asynchronous job outcomes for recovery.
• Human-in-the-loop approval gates prevent generated content from publishing automatically.

Reflection & Engineering Learnings

The platform made a broader lesson concrete: agent complexity must be matched by explicit approval states, durable job records, and observable service boundaries.