A fair question
Couldn’t I just tell ChatGPT?
For one meal, yes
Describe tonight’s dinner to any good AI assistant and you’ll get an answer, and a plausible one. Graham, an AI nutrition coach, isn’t built for tonight. He’s built for the fortieth dinner, the running total, and the week the scale stops moving. That part is bookkeeping, and bookkeeping is the one job a language model shouldn’t do in its head.
The listening is the easy part
Models are good at food. Say “two eggs, a slice of toast and a coffee with milk” and you get the right three things in sensible portions. Graham uses a language model for exactly that, and for talking back to you. What he never hands it is the arithmetic: your totals, your targets and your trend are worked out in code.
What a food log has to get right
The arithmetic. A language model adds the way it writes: fluently, confidently, and usually right. Usually is fine for a sentence. It isn’t fine for a total you’ll check every night for months. We learned that from Graham himself. Early on he’d sometimes announce a day’s total he’d worked out on the way past, sounding exactly as sure as when it was real. So now he can’t. Every total and every “left for today” comes from code, and a figure about your day that didn’t come from somewhere gets caught before it reaches you.
Where a number comes from. Food figures come from USDA data or from a label you read him. When there’s nothing to look up, like someone else’s casserole, he makes an honest call and marks it with a ~, so you can always tell which numbers are firm.
Asking instead of guessing. When he doesn’t know, he asks. “Was that the large?” costs you a second and saves a guess you’d never have noticed.
Your usual, learned once. Teach him your breakfast smoothie and it comes back the same way every time you mention it. Change the recipe and he updates it, and it logs the new way from then on.
A plan that waits for you. Graham keeps your weight trend, not just this morning’s number. Once a week he looks at how it’s moving against your target, and if the target should change he proposes it and waits. Nothing moves until you agree.
Where your food log lives
On your phone. There’s no account to make. To answer you, Graham sends your message and the context he reads it with, like your targets, today’s log and the foods you’ve taught him, through our server to an AI model provider whose terms don’t allow training on it. Apart from feedback you choose to send us, our server doesn’t keep it. The full list is on the privacy page.
Want to try it?
Graham is on the App Store. A free seven-day trial, then $9.99 a month or $59.99 a year.