My daughter has epilepsy. She also has multiple other disabilities, so keeping her safe means a lot more than managing medication. I also have to think about how everything around her might affect her: what she eats, what cleaning products are used in her home, and what toiletries are in the bathroom.
Most people don't know that certain chemical compounds and essential oils can trigger seizures. Sometimes it happens through smell alone, without anything being swallowed. Eucalyptus is one of the best-known examples, and there are plenty of others. For families like mine, the usual method is to crouch in a supermarket aisle and squint at a tiny label. You check it against a list you're trying to hold in your head, and hope you haven't missed anything.
It's exhausting. On the face of it, it's also exactly the kind of problem AI should be able to help with.
Why most "helpful" AI tools aren't.
AI tools built for people with complex health needs often fail in the same way. They're designed by people who don't have those needs, for a user they've imagined rather than met.
The result usually works on paper. It can answer questions and look things up. But it adds friction at exactly the moments when friction hurts most. Picture yourself in a supermarket with not much time and a lot on your mind. You're tired, because managing disability is exhausting. You need a clear answer, fast, and you don't have the energy to decode a response full of probabilities and caveats.
Lived experience changes the product requirements, and the changes go much deeper than the wording or a few added accessibility features. It changes what the product is trying to do in the first place.
What we built.
EPCheck is a free community safety tool that runs in your browser, so there's nothing to download and you don't need an account. You type a product name or upload a photo of the ingredient label. EPCheck checks the ingredients against a database of known and suspected seizure triggers and gives you a result within seconds.
Each flagged ingredient comes with context: the risk level, how exposure happens (topical, inhalation or ingestion), what the evidence shows, which common products contain it, safer alternatives, and how quickly a reaction might happen. The aim is to give you enough to make a decision, right there in the aisle.
The design decisions that came from living it.
Speed over completeness. In a supermarket, a first result in two seconds, with the option to dig deeper, is more useful than a full report in thirty. We built for how the tool gets used, and left out anything that was only there to show off what it could do.
No account required. Making people register before they can use a safety tool puts a barrier in their way right when they most need help. EPCheck works straight away, for anyone, with no sign-up. That one was never up for debate.
Context with every flag. If you tell someone "this ingredient is risky" without saying why, how risky, or how exposure happens, you leave them anxious with nothing they can do about it. We give people enough information to make their own decision.
Community knowledge counts. Peer-reviewed research is the foundation. But the epilepsy community has built up a huge amount of practical knowledge about what triggers seizures, and much of it hasn't made it into published research. EPCheck takes community feedback seriously and treats it as evidence in its own right.
What this means for how you build AI tools.
EPCheck is one example of a broader idea. The most useful AI tools for communities with complex, high-stakes needs are the ones built closest to the people who'll use them. That doesn't mean every AI tool has to be built by someone with the condition it serves. It means the people who'll use it need a say in what it should do.
It means designing for the situation people are really in. For a lot of accessibility and health tools, that's a heavy mental load, high emotional stakes and not much time, which looks nothing like the polished demo. It means treating community knowledge as real input. And it means being willing to limit what the tool does. The pull in AI product development is always to add more. I think the harder and more useful skill is knowing what to leave out.