Health-critical AI

FoodSafe

A label scanner built for one person with 74 confirmed food intolerances. It reads the ingredients, checks them against her profile, and gives a clear verdict — or tells her it couldn't be sure.

Status
Live in production
Built for
One named person
Profile size
74 triggers
My role
Design, build, run
Try it · tap a product
FoodSafeScan
Her profile
74 triggers · she maintains it, not us
Scan a product
Ready

Pick a product above and FoodSafe will check its ingredients against all 74 triggers.

The problem

Someone close to me has seventy-four separate food intolerances, confirmed by testing. Shopping means holding all of them in your head at once while reading small print in a cold aisle with someone waiting behind you. Get it wrong and the cost isn't inconvenience — it's days of being unwell.

She was doing this with a folded paper list. It worked, slowly, until a product changed its recipe or an ingredient hid behind a name she didn't recognise.

The constraint that shaped everything

This is the only thing I've built where a wrong answer has a physical cost. That single fact changed how the whole application is written.

A tool that is confidently wrong about food is worse than no tool at all, because she would stop reading the label herself.

So the scoring logic — the part that decides safe, caution or avoid — is the piece I treat as untouchable. It's plain, readable rules rather than a model improvising, it's tested, and no change goes near it casually. The AI in FoodSafe reads and normalises messy ingredient text. It does not get a vote on the verdict.

What I built

She scans a product. The app reads the ingredient list, matches every item against her profile, and returns one of three answers — not two. Safe, check with care, do not eat. The middle one matters: plenty of products aren't dangerous but aren't clean either, and collapsing that into "safe" would be a lie.

Whatever it flags, it names. The verdict always lists which ingredients matched and how severe each one is on her profile, so she can disagree with it. She maintains her own trigger list; I don't edit it for her.

When it refuses

If the label is torn, blurred or ambiguous, FoodSafe gives no verdict at all. It says it couldn't read the product and stops.

That was a deliberate decision and it's the one I'd defend hardest. A system that always answers teaches you to trust it; a system that sometimes says "I don't know" teaches you when to check. The second one is the one you want standing between a person and their dinner.

Where it stands

Live, and in use by the person it was built for. It's the oldest thing on this page and the one that set the standard for everything after it: when a system can't verify something, it stops and says so.

Free first consultation

You've seen the software. Let's talk about yours.

One conversation, no cost, no obligation — and an honest answer about whether there's something here worth building.

Book a free consultation →