Ertas for Content Creation

The Challenge

Content teams at media companies, agencies, and brand marketing departments face a relentless demand for high-quality, on-brand content across an ever-growing number of channels: blog posts, social media, email campaigns, product pages, video scripts, and more. Generic AI writing tools can produce grammatically correct text quickly, but the output reads like it was written by the same anonymous assistant powering every other brand. It lacks the distinctive voice, editorial standards, and institutional knowledge that separate great content from filler.

The problem compounds at scale. A media company publishing hundreds of articles per week or a brand maintaining content across dozens of product lines cannot manually edit every AI draft into compliance with style guides. Prompt engineering helps at the margins, but it cannot teach a model the nuanced voice differences between a luxury brand's Instagram captions and its investor communications. Teams end up spending almost as much time fixing AI-generated content as they would writing from scratch, negating the productivity promise entirely.

The Solution

Ertas lets content teams train AI models built for their brand. In Ertas Studio, editorial and marketing teams fine-tune open models on datasets of approved past content (published articles, high-performing social posts, email campaigns and product copy), each paired with the brief that produced it. In Data Craft, Studio's guided dataset builder, editors review and edit examples and mark the ones that show the brand at its best as approved. Teams can train a separate model for each voice, one for the corporate blog, another for social media, a third for technical documentation, and run them in parallel to compare. Each model picks up the vocabulary, sentence rhythm and editorial preferences in its examples.

Trained models appear in Hub, where the team downloads them as GGUF for Ollama, llama.cpp or LM Studio, or as full safetensors for vLLM. They run on the company's own servers or a writer's laptop, and connect to a CMS or editorial tool through a small integration the team builds against that runtime. Writers get first drafts that already sound close to the brand, which shortens the editing cycle. When the voice evolves, editors add newly approved examples and the team retrains. The models belong to the brand and need nothing from Ertas at runtime.

Key Features

Example Workflow

A direct-to-consumer skincare brand publishes 40 blog articles, 200 social media posts and 15 email campaigns per month across three product lines, each with a slightly different voice. The marketing team exports two years of approved content, tagged by product line, channel and content type, and pairs each piece with the brief it was written from to make a JSONL dataset. Editors review a sample in Data Craft and fix examples that no longer match the current guidelines. In Ertas Studio, the team fine-tunes three models from a Mistral 7B base in parallel: one for the clinical-authority voice of the professional line, one for the playful tone of the everyday products, and one for the aspirational feel of the premium range. They download all three as GGUF and run them in Ollama on an office server, connected to the brand's Contentful workspace through a small custom plugin. Writers pick the product line before drafting, and the matching model writes the first draft. Before rolling the workflow out to every channel, the editorial team spends a month comparing AI-assisted drafts with its usual work on editing time and its internal brand-consistency rubric.