I work for a company that offers nutrition tracking on an app in the App Store.
We are not shipping camera functionality yet. But our concept is to not necessarily guarantee the accuracy of portions but to make lookup easier.
We also spent the time to get the AI integrated with a verified database. This made our results far more accurate.
We tended to find that without the lookup the calories and macros would be generally correct. The math was usually within a margin of error of 5%. This was acceptable except that… there was no micronutrient values and you couldn’t really adjust the portions at all. The system just dumps the macros and while you can halve something… the user experience isn’t great.
Ultimately, if you want precision: manual entry is the only way to go. I feel like out approach will end up being very great once we work out the kinks. Our search isn’t spectacular and as a team we are learning a lot of about prompt engineering and how to make best use of the AI.
Yes, I would think that would work better indeed. As a augmentation or help tool. I would love to be able to say to MyFitnessPal that 'I ate this and that food, same as usual, and oh yeah drank this.' Just as a easier input interface. I wouldn't trust a pure AI solution without some proper database behind it.
Yeah. The big problem is that "augmentation" is hard because we (humans) have an internal process for how we think about things that is hard to define and building a flowchart for how we understand foods doesn't even necessarily capture things. Very well. You can take something like "chocolate chip pancakes" where the context can be "<brand> <item>" or "<modifier> <food item>". And then you can search.
But even though we've integrated it with a good food database, the process of searching isn't great because sometimes things like brand names don't get recognized and/or modifiers may get confused because... is it a brand? Is it a way of preparing something?
Ultimately we are working on improving how our search works by not just searching by the name, but by getting information about brand, the product, and possible serving options as well. These would better inform the search and allow us to, say, fallback without a brand if we can't find the brand.
The other problem has to do with variant detection. I can say "kirkland sous vide egg bites" but there are 3-4 variants of them. And right now most databases are just "here is the item you requested" without looking at possible variants, which is a problem that we are going to end up solving ourselves.
It's been interesting because we've learned a lot about how people "think" it should work vs. how it actually works.
Does that work for homemade food as well? The vast majority of the food we eat is homemade with recipes that don’t have any sort of nutritional information. I’ve always wished there was a simple way to figure out the calories. Taking a picture would be ideal.
For homemade food it should be easier to make reliable estimates of the calorie content, because you know with certainty all the food ingredients and their amounts.
The food ingredients with the highest calorie content, like various kinds of seeds or nuts or flour or meal or oil or fat or sugar or dried fruits or dairy products, come usually with calorie estimates from their vendors.
For other ingredients, like various kinds of meat or of fresh vegetables or fruits, there are online databases with typical nutritional information, like the USDA database. Some of that information can even be found in the corresponding Wikipedia pages.
Weighing everything (rather than using volumetric measures) is generally going to be the BEST way to ensure consistency and accuracy.
What's also important is that, in general, even if you are 20% off on something (e.g. I logged 2200 calories but I actually consumed 2600 calories) AND you are planning to eat at a caloric deficit, this usually will mean that you will still lose weight or body recomp. It'll just take a little more time.
But if you are just not tracking, it's _so easy_ to miscalculate your intake to the point where you think "oh this isn't that bad." However, the truth is you consumed 4200 calories and that's a big surplus.
So I/we tend to find the value partially in "simple tracking" to get you aware of what you are actually consuming and then find that transitioning to specific portions to be helpful for "dialing in" and achieving specific targets/goals.
We are not shipping camera functionality yet. But our concept is to not necessarily guarantee the accuracy of portions but to make lookup easier.
We also spent the time to get the AI integrated with a verified database. This made our results far more accurate.
We tended to find that without the lookup the calories and macros would be generally correct. The math was usually within a margin of error of 5%. This was acceptable except that… there was no micronutrient values and you couldn’t really adjust the portions at all. The system just dumps the macros and while you can halve something… the user experience isn’t great.
Ultimately, if you want precision: manual entry is the only way to go. I feel like out approach will end up being very great once we work out the kinks. Our search isn’t spectacular and as a team we are learning a lot of about prompt engineering and how to make best use of the AI.