Want something like this?
Tell us what yours would need to do differently. We reply with a scope and a price.
Work · 06 · Case study
Social publishing pipeline
An automated publishing pipeline in which an AI drafts every post and paints every picture, and is never allowed to publish either.
The problem
Posting is a chore shaped in the worst possible way: four platforms, an angle to find, an image to make, the same thought rewritten three ways. Small teams either stop posting or hand the account to a tool that posts by itself, and an LLM with write access to a company page is a liability on a timer.
How it works
A schedule, a Drive folder or a Telegram command starts a run. The workflow picks a topic from a rotation that remembers what it said last time, one model call returns four platform drafts, a validator treats the result as untrusted input, and a second model call art-directs a poster for the post. Everything lands in a Telegram chat as one message with two buttons. Nothing is published until a person taps Approve; the publish credentials live on the far side of that tap.
The pictures
Every post gets its own poster: a model reads the finished text and proposes one visual concept, code assembles the prompt from four rotating axes with prime counts (1,155 combinations before an exact repeat), and the organisation’s logo is stamped on afterwards by code, not by the model. The model paints before the tap; code stamps after it.
What breaks, and what happens then
Everything degrades, nothing halts, is the rule. The picture side lives up to it: art director down, plainer prompt; image model down, the branded card; logo step down, the picture ships unstamped. The rest is catching up, and the table says so.
Built with
Self-hosted n8n, Gemini for text, OpenAI for images, Postgres for memory, Telegram for approval, the LinkedIn and Meta APIs for publishing. Running for the lab and for a real-estate client, from one codebase with a different brand each. The full write-up, including the outages, is on its way to the field notes.