An automated newsletter that digs through the world's public archives and comes back, every other day, with one remarkable image and five minutes of properly researched writing about it.
The cost design
The pipeline is a funnel. Discovery and filtering cost nothing; the model is only called on candidates that already survived cheaper checks — and the biggest calls happen once, on a single winner.
Architecture
If a draft fails quality control it is rewritten once, then abandoned for the runner-up. If nothing clears the bar, there is simply no edition that day.
Trust & safety
Key design decisions
Length, attribution, sourcing depth, licence and safety are checked in code and block publication. The model's own review can lower a score, but never raise one past a mechanical veto.
Research reads Wikipedia for orientation, then follows its references to the museums, agencies and journals that hold the evidence. Wikipedia alone can't support a key claim.
Source quality and image quality are computed from the pipeline's own data — a model is a poor judge of its own evidence.
"Is this basically Tuesday's story?" is answered with token overlap on titles and entities: free, dependency-free and good enough — swappable later if it isn't.
Eleven archives and several research pages per candidate are fetched concurrently, turning a fifteen-minute run into about two.
pydantic, httpx, Jinja2, PyYAML and python-dotenv. No ORM for a nine-column table, no web framework when there is nothing to serve.
Happy to walk through the architecture and trade-offs.