← Back to all work

cravingtoolkit.com — RAG and AI content pipeline

The problem

A recovery-focused content site needs to publish high-quality, safety-conscious content at scale and respond to user queries with accurate, sourced information drawn from a curated knowledge base. YMYL (Your Money or Your Life) constraints mean every output needs editorial guardrails.

What I built

An automated SEO publishing pipeline using LLM-generated drafts with editorial guardrails, plus a RAG-powered Q&A layer over the site's own content and vetted external sources.

Stack

Python, OpenAI and Anthropic APIs, vector embeddings, automated content workflows, structured editorial review.

What I learned

• Safety-conscious AI in sensitive domains demands more than a system prompt — refusal patterns, retrieval-only answer modes, and explicit escalation paths matter more than the choice of base model.
• Content quality thresholds keep rising. Google's helpful-content updates and improving AI-content detection mean LLM output has to be edited, sourced, and structurally distinguishable from generic generations to earn and keep ranking.

Related work

• Order Management Systems at dwpbank — production Spring Boot, Kafka, Angular at financial-services scale.
• Multi-agent automation infrastructure — self-hosted LLM agents with provider routing and cost discipline.