Our summary
Choosing coffee can mean sorting through thousands of bags with different tastes, prices and roast levels. Sam French describes a website where a reader writes one sentence about what they want. Jev interprets those preferences and helps choose coffees from a smaller list prepared by the site's code.
The first version guessed a budget even when the reader never mentioned money. French reports adding a question about whether a budget was stated. He describes database rules enforcing requirements such as decaf before ranking: an earlier test confidently selected ordinary coffee from a list containing no decaf.
If the model fails or takes too long, the site falls back to keyword matching and labels it that way. French's timings and costs come from small reported runs, not an independent test. The reader's words and candidate descriptions go to hosted AI services, so this is not private, on-device matching.
Key takeaways
- Ask whether a preference was stated before using a guessed preference to remove options.
- Enforce firm requirements before asking AI to rank a list. A confident winner can still be unsuitable if every option is wrong.
- Label the backup method honestly and test database rules on the database used in production. A substitute test database missed real failures here.