Still confused? Let’s talk
exclude: nuts
Guide · 7 min read
When an AI agent answers "is this nut-free?" or "is this safe in pregnancy?", the data behind it decides whether someone gets hurt. Five rules for building data an agent can be trusted with.
Guide · 6 min read
Restaurant and grocery chatbots get allergen questions every day, and menu data was never written to answer them. Where the errors come from, and how to structure allergen data an agent can rely on.
Guide · 5 min read
An agent's answer is only as trustworthy as the data under it, and that data needs a trail. What a useful receipt holds, and why "the model said so" isn't one.
Field note · 5 min read
A re-sync can drop a safety value from a hundred records without anyone deciding it should. How each new version is compared with the last release before it goes out.
Field note · 6 min read
A prompt, model or rule change can make a catalog better overall and worse exactly where it matters. How a golden set catches the release that loses a safety value.
Field note · 5 min read
Teams want their own model for cost, contracts or data residency. What stays enforced whichever model answers, and the one thing we disclose instead of solving.
Field note · 4 min read
An answer key says the values are right. It doesn't say the search an agent sends still answers right. Scenarios check that, and hold back a release that would break one.
Field note · 4 min read
Similar-sounding exercises kept ranking next to each other. The fix was to stop letting names speak for meaning.
Merged, not replaced. Dairy stays.
Field note · 5 min read
How a shared, cross-restaurant dish catalog saves repeated AI work, and why its allergen data is only ever allowed to grow.
Field note · 5 min read
An AI coach should never recommend an exercise that hurts someone. Here's how to enforce that without trusting the model to remember.
2 settled by rules, 1 model call.
Field note · 4 min read
The cheapest, most testable enrichment step is the one that never calls a model. Here's where we draw that line.
19 / 19 agent scenarios passing
Field note · 4 min read
Agents regress quietly. Scripted scenarios, repeatable golden cases and observable tools are how we catch it.
Case study · 6 min read
1,324 exercises enriched and embedded, with safety enforced end to end.
Case study · 6 min read
Any menu format in, 14 allergen classes tracked, and a shared catalog so no dish is enriched twice.