Skip to content

Akros · Accuracy

We measure ourselves in public.

Every week our benchmark runner scores Akros AI food-photo estimates against weighed-truth meals from public labelled datasets. We do not run head-to-head comparisons against competitors that do not publish per-meal data; we cite their published ranges instead.

This week’s Akros runs

Benchmarks publish weekly after launch.

Our nightly Cloudflare cron seeds this table with results from the public Snap-It and Recipe1M test sets. The first run lands within seven days of public launch.

Published competitor ranges

We do not fabricate Cal AI vs Akros per-meal comparisons because Cal AI does not expose per-meal accuracy data. Below are the figures each competitor (or independent reviewer) has published — sourced and linked.

How we run the benchmark

A Cloudflare Cron Trigger fires nightly at 04:00 UTC. A Durable Object loads the current dataset (Snap-It or Recipe1M), pipes each labelled image through the same food-photo endpoint your iPhone calls, records the returned macros, and computes mean absolute error against the labelled truth. Median latency and median per-meal cost are stored alongside MAE.

Results go into a public Postgres table read by this page. We refuse to scrub failed runs. If a release regresses accuracy, the next week’s row makes it visible to you before we tell ourselves it’s fine.

Full methodology — dataset licence, exclusion rules, confidence intervals, vendor comparison limits — lives at akros.life/accuracy/methodology.

Akros is a personal wellness app. It is not a medical device, does not provide medical advice, and is not a substitute for consultation with a licensed clinician.