Resources

AI pipeline pre-flight checklist

Run this before you point your pipeline at real traffic. Every item on it is something I have personally skipped and regretted.

• The tenant boundary is enforced in code, not by convention — a request without a tenant id is rejected before any query runs.
• Every model call has a ceiling: max tokens, max retries, and a timeout that is actually enforced.
• Retrieval runs before generation, so the prompt carries the source material instead of the model's memory of it.
• A cheap model checks the output against the topic and brand rules before it reaches the publish step.
• Every prompt, model version and parameter set is recorded, so a bad output can be traced back to the input that produced it.
• Failed publications retry with backoff, and the retry count is capped — a broken selector does not loop forever.
• The publish step is idempotent, or it has a dedup key, so a retry cannot publish the same article twice.
• Token spend per article has a budget, and exceeding it fails the job loudly instead of quietly billing you.

Evaluation worksheet

Fill this in before you change a prompt or a model. If a row has a blank in the "before" column, you do not have a baseline yet — and without one you cannot tell whether the change helped.

Cost worksheet

The one number that decides whether a pipeline is viable is cost per published article. Fill in the middle column with real figures from your own logs — not the vendor's headline price.