Most food apps make you search, scroll, pick a serving and set a meal — for every item, every time. Tally does that once and then remembers.
One log a day on a demo account. No card, no trial timer.
“two eggs and a fairlife”
Numbers from USDA FoodData Central — the food, the portion and the meal already chosen. One tap files it.
Every commit teaches it the phrases you use, the foods they mean and the portions you take. A meal you have logged before resolves from your own history with no AI call at all — one database round-trip. Only a food you have never logged touches the network, once.
Text and voice never estimate. If a food cannot be found, the app says so instead of guessing. The one exception is a photo, where amounts have to be estimated — and there every estimate is flagged and refused at commit until you confirm it.
Accounts are kept apart by Postgres row-level security, not by application code — the app runs as a database role that cannot bypass it. Photos are never stored.
Say “two eggs and a fairlife”. Both rows come back filled in — food, portion and meal.
Vision names what is on the plate and estimates the amounts. Every estimate is flagged and has to be confirmed.
Scan the label. The numbers come from the manufacturer’s own record, not an estimate.
The boring escape hatch, on purpose. Your own foods first, then the USDA database.
Tally is MIT-licensed and built to be run by one person on free tiers. The design document, the phased build plan and the ledger of every review finding are all in the repo.