Content Automation Platform

The problem

Multi-location clients need blogs, service pages, and city pages at a volume no manual writer can sustain, and every new content type used to mean an engineering ticket. Worse, generic AI copy doesn’t sound like the client — it sounds like the model.

What I built

A config-driven content platform: new pipelines — blog posts, service pages, city pages — were added via JSON configuration, not code, so onboarding a new content type didn’t need a redeploy. Generation ran through a unified multi-provider LLM gateway across Claude Sonnet 4 and GPT-5, an LLM guardrail layer enforced per-location business facts as hard constraints before anything generated, and every call was logged to Braintrust for eval and regression monitoring.

Decisions worth noting

• Closed-loop learning from edit diffs. The system diffed what a human edited against what the model generated over a rolling 30-day window, mined the pattern into confidence-scored per-client style rules, and fed those rules back into an automated QA agent. The idea: the system improves from disagreement, every edit is a signal, not just a correction.
• Unified multi-provider gateway with per-call cost attribution. Every generation call routed through one gateway across Claude and GPT-5, with cost tracked per call, not estimated after the fact.
• Prompt-caching economics modeled at real prices. Cache-write and cache-read pricing are different numbers, and the gateway modeled them as what they are — the difference between a cost estimate and a cost the business can actually plan around.

What happened

It didn’t stick in production. The machinery — closed-loop learning from edit diffs, a multi-provider gateway, a config-driven pipeline system — was more platform than the actual content workflow needed. Alyne pivoted to a simpler approach: Claude Skills doing the content generation directly, with far less infrastructure to build and maintain. The lesson outlasted the platform: build the complex thing when the problem genuinely demands it, but don’t mistake “impressive to build” for “right for the job.” I’d rather ship this honestly — built the complex thing, learned from it, chose the simpler thing — than pretend it’s still running.

My role

Sole architect and engineer at Alyne — backend, frontend, and the LLM integration. Source is private, so there’s no repo link here. The platform itself is retired; the write-up is what’s left, as a lessons case study.