The job was plain: keep three of our own brands posting on social without a person at a keyboard every morning. The easy version is one big writing model and a publish button. We have all seen what that produces. The model writes something that reads well, makes up a project that never happened to make the point land, and posts it under your name before anyone notices.
So the pipeline we built is mostly about what happens between the model writing a draft and a human seeing it. Through the night, a machine quietly does the rounds — gather material, summarise it, write a draft, check it, turn it into an image, line it up to post. The part that writes is a large, capable model. It is also the source of the problem, because the better a model writes, the more confidently it makes things up.
Why one check is not enough
The instinct is to add a second model to review the first one, and call it done. That fails in a specific way. A single reviewer is either too cheap to catch the dangerous mistakes, or too slow to run on every draft. You end up tuning one model to do two jobs it is bad at doing at once.
The two ways a draft can be wrong are genuinely different problems.
- Wrong voice. The draft sounds off for the brand — too loud, or too corporate where it should be plain. This is cheap to spot and it happens often. You want it caught in a blink.
- Made-up facts. The draft says something false — a result we never got, a client we never had, a number nobody measured. This is rare, harder to spot, and the only one that can actually hurt you. You want it caught by something that takes its time.
Putting both on one model forces a bad trade. So we put them one after the other instead.
The arrangement
A smaller, faster checker runs first. Its only job is voice. It reads the draft against the brand's rules and throws back anything that does not sound right. Most drafts that fail, fail here, and they fail quickly. It is cheap, so we can afford to run it on everything.
Whatever survives goes to the slow gate: a careful review step whose whole job is to fact-check the big model's claims and rewrite the examples it invented. This step is the expensive one. It is also the one that earns its keep, because it is the only thing standing between a confident piece of fiction and your audience.
The fast check protects the brand's voice. The slow check protects the brand's word. Only one of those is recoverable if you get it wrong.
Only after both checks does a draft reach a person to approve. By then it sounds right and it is true, and the person reviewing is making an editorial call — not cleaning up a mess.
The bit worth keeping
Everything else around this is ordinary plumbing: turning a draft into a branded image, scheduling it to post, recycling older approved posts when the queue runs dry, a switch to pause and resume the whole thing with one message. None of that is the lesson.
The lesson is the order. A cheap, fast check for the common cheap mistake. A slow, careful check for the rare expensive one. A human last. Collapse those two checks into one model and you will publish a confident lie eventually. The cost of running them in sequence is a few seconds and one extra call. The cost of not is your name on something that is not true.
Under the hood
The pipeline runs on a single headless workstation as eight scheduled cron jobs through the night — collect, summarise, generate, validate, render, queue, publish.
The generator is a large model, around 253 billion parameters. Strong writer; also the source of the fabrication risk.
The fast check (voice) is a smaller validator, around 120 billion parameters but a mixture-of-experts design, so only a fraction (~12B) is active per token. That makes it cheap enough to run on every draft. It scores the draft against the brand spec and rejects on tone alone.
The slow gate (facts) is a review agent driven by an authoritative prompt. Its job is to fact-check the generator's claims and rewrite any examples it invented, rather than just flag them. This is the expensive call, run only on drafts that already passed the voice check.
After both, the draft enters a human approval queue — in practice a messaging channel where a person taps to approve.
The surrounding machinery: visuals are rendered from HTML templates and screenshotted with headless Chrome; AI-generated backgrounds are blended only into the dark regions of a card so text stays legible; publishing goes out through a scheduling API; an evergreen pool recycles approved posts when the queue empties; and a single message pauses or resumes the run. Status telemetry — queue depth and approval rate — lands in a time-series database so the pipeline can be monitored without logging into the box.