Meal logging

Goal: a health or coaching agent sees a meal photo and logs structured nutrition - without trusting a single vision-model guess.

Estimate (calibrated, self-disclosing)

POST /v1/meal-vision   $0.05/image
input:  {"image_url": "...", "hints": "optional cuisine context"}
output: items[] (dish, portion, kcal, P/F/C, confidence),
        kcal_range (two-sided, calibrated), consensus spread + agreement

Every call runs a 3-run consensus and reports the median run, the per-call kcal spread and item agreement - reproducibility your agent can probe directly. Pass a cuisine hint for regional home-style meals (measured accuracy gain).

Why the numbers can be trusted

Calibrated against 30 dietitian-portioned weighed meals; the disclosed kcal range contained the true value in 26 of 30. Method and systematic error classes are published at /vision-calibration. Not for allergen detection or medical decisions.

Try before paying

GET /samples/meal-vision   free recorded response shape