Content Engine: AI content automation
A production B2B AI content pipeline built for high-volume automated publishing. It ingests sources, drafts with LLMs, routes through human review, and automates content distribution to X, LinkedIn, and Facebook, running daily.
Discuss this project
Runs every day in production, automated end to end with a human approval gate.
Publishing at the speed of the news
In a fast-moving news niche, the team had to publish a high volume of content every single day, and be early, not just accurate. Sources had to be watched constantly, drafts written and re-written, posts scheduled, and the same stories pushed out across multiple social channels by hand.
That meant long days of repetitive work, a constant risk of missing a story, and no realistic way to scale output without simply hiring more people. The bottleneck wasn’t ideas; it was the manual machinery between a source and a published, distributed post.
How it works
I built Content Engine as an end-to-end pipeline that runs on a schedule, every day.
1. Ingest: It continuously pulls from the team’s sources, deduplicates, and normalizes everything into a single queue of candidate items.
2. Draft (LLM): Large language models turn raw material into formatted drafts, guided by prompts and guardrails tuned to the team’s voice and format.
3. Human review: Nothing ships unattended. A person approves, edits, or rejects each draft; rejected items loop back rather than going out.
4. Publish: Approved content is published to the primary destination automatically.
5. Auto-distribute: The same approved item is fanned out to X, LinkedIn, and Facebook, formatted per channel.
Human in the loop, by design
Automation handles the volume; people keep the standards. The review step is the heart of the system: every piece passes a human gate before publication, so the team keeps full editorial control while the machinery does the repetitive lifting.
The pipeline suggests, formats, and queues; it never publishes something a person hasn’t signed off on. That’s what makes it safe to run at high volume in a niche where being wrong in public is expensive.
Stack
Python · LLM drafting with prompt caching (Anthropic Claude / OpenAI) · Perplexity API for fact-sourcing · human review queue · automated publishing via CMS API · OG-image generation · structured data and SEO schema · per-channel formatting with scheduled distribution · cron-based orchestration.
Results
• Runs daily in production: the pipeline has operated on a daily schedule since 2025, end to end, without manual babysitting.
• One approval, every channel: a single editorial decision publishes the piece and fans it out to X, LinkedIn, and Facebook, formatted per channel.
• Manual machinery removed: sourcing, drafting, formatting, scheduling, and cross-posting are automated; the only human step left is editorial judgment.
The operation's traffic and volume figures belong to the product, so I keep them off this page. I'm happy to walk through them in a call.
Project outcomes
• Scale without headcount: A small team publishes at a volume that previously would have required hiring, because the repetitive work is automated end to end.
• Consistent voice: LLM drafting plus human review keeps every post on-format and on-brand.
• One source, every channel: Approving an item once distributes it everywhere, so nothing is re-typed or forgotten.
• Editorial control retained: The human gate means the team never trades quality or judgment for speed.