"Can I eat this?" is one of the most frequent questions during pregnancy, and it rarely comes up at home. It comes up at a restaurant table, with a menu in hand, a waiter waiting and no time to search through official guidelines. That is the exact moment SafeBloom, the app designed and built by Vivig Consulting, was made for. Its core feature, the AI Menu Scanner, is also a good case study for any company wondering how to turn generative AI into a product people can trust.
What the Menu Scanner does
The flow is deliberately short. You photograph the restaurant menu, printed, on a chalkboard or, since version 1.4, a digital menu reached through a QR code. Within seconds every dish gets a traffic light:
- Green: safe.
- Yellow: safe with precautions, and the app tells you what to ask the waiter (for example whether the eggs in a tiramisù are cooked or the mozzarella is pasteurised).
- Red: better avoided, with alternatives from the same menu.
Each verdict comes with the reason and the institutional sources behind it, such as the Italian National Institute of Health (ISS) and the European Food Safety Authority (EFSA). The analysis looks for the real risk factors: raw or undercooked foods, unpasteurised products, alcohol, mercury, listeria.
Five design choices that make AI trustworthy
Reading a menu is easy for a modern AI model. Making the answer reliable enough for a pregnant woman to act on is the hard part. These are the choices that made the difference, and they apply well beyond food.
1. A human expert sets the rules
Recipes, food sheets and the Menu Scanner's classification criteria are validated by a nutritionist specialised in pregnancy and breastfeeding, with a scientific board behind the project. The AI applies the criteria; it does not invent them. For any business, this is the most important lesson: the model brings speed, the expert brings the standard.
2. Every answer shows its "why"
A bare "no" creates anxiety; a "no, because it is raw fish, ask for the cooked version" creates a decision. Explaining the reason and citing the source turns an opaque prediction into advice the user can check.
3. Context changes the answer
The user profile includes the week of pregnancy or breastfeeding, the toxoplasmosis test result, gestational diabetes and allergies, and the results adapt accordingly. The same dish can be green for one person and yellow for another. Personalisation is not a gimmick here: it is what makes the output correct.
4. Clear limits, stated up front
SafeBloom says plainly that it provides information based on institutional sources and does not replace a doctor's advice, and that for non-certified restaurants the analysis is based only on the menu text, so staff should always be asked. Declaring what the AI cannot know is part of the product, not small print.
5. Privacy as a feature
Health-related profile data stays private and can be deleted by the user at any time. In Europe, under the GDPR and with the AI Act now in force, this is not optional, and users notice it.
From one app to an ecosystem
The scanner is the entry point of a wider system. Alongside it, Sofia, the AI nutrition assistant, answers questions about pregnancy and breastfeeding in Italian and English. A map shows more than 150,000 places in Italy and Spain, including SafeBloom-certified restaurants: venues whose staff completed an 8-hour training on food safety in pregnancy, with a final exam. Restaurants get their own dashboard for certification, training, marketing materials and reviews. The app and website are available in five languages, on both iOS and Android.
The model is freemium: the free plan includes 10 menu scans and 15 questions to Sofia per month, plus restaurant search, recipes and articles, while the Premium plan unlocks unlimited use and a consultation with the nutritionist.
What businesses can take from it
- Start from a precise moment of need, not from the technology. "At the table, menu in hand" is more useful than "an AI for nutrition".
- Keep the expert in the loop and make the AI apply documented, verifiable criteria.
- Design the output for action: a traffic light, a reason, an alternative, a question to ask.
- Write the limits into the interface, especially in regulated or sensitive areas.
- Treat data protection as part of the value, not as a compliance afterthought.
How Vivig can help
SafeBloom is one of the apps Vivig Consulting has designed, built and published end to end, together with NettoChef and SafeGuard: you can see them all on our Projects page. If your team wants to learn how to apply the same approach to its own processes, our AI Evolution Program trains companies to use AI in practice, safely and with measurable results. Have an idea for an app or an AI workflow? Let's talk.