Building FitScan AI (meal photo → macros)

Building FitScan AI (meal photo → macros) and the part that keeps surprising me isn’t the first AI guess — it’s the correction UX after it.

Photo logging is great until the model thinks your “light olive oil drizzle” is a tablespoon, or merges two sides into one item. If fixing that takes more taps than searching MyFitnessPal, people abandon the scan and go back to typing.

What’s been working for us:

• Show calories + P/C/F immediately, but treat it as editable draft state

• Let people tap one food on the plate and nudge portion / swap the item without redoing the whole scan

• Keep a one-tap “looks right → log” path for clean plates so good days stay fast

• Fall back to short text describe when the photo is messy (dark restaurant lighting, covered bowls)

Accuracy without easy correction is theater. Correction without speed is just a slower diary.

If you’re into this space or want to try the flow: fitscanai.org