A content engine that compounds
The problem
A content-heavy agency was hitting a real ceiling. Demand from their clients was rising fast. The editorial team couldn't grow at the same pace. And the off-the-shelf AI tools they'd tried produced output that wouldn't survive review, which meant editors were rewriting drafts from scratch instead of editing them. The math broke either way. They needed to scale production without scaling the team, and they needed quality to compound from one batch to the next, not regress.
The engagement
Build a content production engine that ships publication-ready articles on the first pass, gets materially better with every editorial cycle, and serves international clients without doubling editorial work.
What we did
We built a multi-stage generation pipeline that orchestrates several LLMs across the workflow, each doing the work it does best. Drafting, voice enforcement, citation verification, external-link validation, and format checks run as discrete stages with quality gates between them. Output doesn't move forward until it passes. Bad articles never reach an editor.
The pipeline treats editorial feedback as a first-class input. Every issue caught in review becomes a new automatic guardrail in the next cycle: voice deviations, structural gaps, off-topic links, weak claims. The system is materially better today than six months ago, not by retraining a model but by continuously tightening the operating context around it. The agency's editors are effectively training the pipeline by doing their normal review work.
On top of the core pipeline, we shipped a multi-language extension (slug, meta, and body all exported in the target language) and a batch interface that takes CSV in and produces production-ready articles out, with full tracking and resumability on every job.
The impact
211 publication-ready articles produced in the most recent batch. Editor time per article down ~65% and still dropping as the guardrails mature. Six languages now supported through the translation pathway. Nine documented failure modes from editor feedback are built into the pipeline as automatic checks. Editorial throughput scaled past what the team could have reached alone, and quality is compounding from cycle to cycle rather than regressing.