Skip to content
Pantrly

How the AI works

Model-agnostic, and honest about it

Pantrly does not exist without modern language and vision models. Here is exactly which model does what, how allergies are enforced twice, and what happens when a lab has a bad day.
01recognize

Read what you have

If you snap a photo, a vision model reads the food it can genuinely see into a plain list. It is told to name only what it can make out and to leave off anything hidden or too blurry, because a confident wrong item is worse than a missing one.

02cook

Suggest real meals

Your ingredients, your diet, your allergies and your time go to a balanced language model with a prompt that asks for real dishes with real steps and honest times, built to lean on what you already have. It is asked for three varied meals, best-fit first.

03cook

Use up what is going bad, and shop for the rest

The same step is told which ingredients are about to turn and builds around them first, with a plain use-it-up note. The shopping list is assembled in code from only the items no recipe could cover from your pantry.

04plan

Plan a week when you ask

A whole week of dinners from one pantry is a bigger reasoning job, so it escalates to a frontier model. Small clean-up jobs, like tidying a messy typed list, go to a cheap fast model instead.

The routing table

Each task goes to the model best at it

This table is generated from the same file the app routes through, so it cannot drift from what actually runs. An eye marks a step that reads an image.

TaskModelProvider
Turn your ingredients into real mealsgpt-4.1-miniopenai
Read a photo of your fridge into a listgpt-4.1-miniopenai
Plan a week of dinners from your pantrygpt-4.1openai
Clean up a messy ingredient listgpt-4.1-nanoopenai

The allergy guardrail

Enforced twice

The model is told allergies are absolute. Then, after it answers, the code checks every recipe against an expanded list of allergen words, so butter is caught as dairy and tahini as sesame, and any recipe that mentions one is dropped before you see it.

Escalation

Bigger jobs, bigger model

A quick weeknight suggestion runs on the balanced model. A long request, like a full week planned from one pantry, escalates to a frontier model automatically. The threshold is a configuration value, not a code change.

Fallback

One lab down is not your problem

If the first model is slow or unavailable, the request walks a fallback chain to the next capable model, and for a photo the chain only includes models that can read an image. You get dinner either way.

Candidates

Declared behind one interface

Every model Pantrly can route to is declared with its real name. The ones from other labs switch on the moment their key is present.

gpt-4.1

openai · frontier

a week of dinners planned at once

gpt-4.1-mini

openai · balanced

real recipes from your ingredients

gpt-4.1-nano

openai · fast

cleaning up a messy typed list

claude-sonnet

anthropic · frontier

careful recipes for tricky diets

gemini-flash

google · fast

fast photo reading

llama-open

meta · open

self-hosted suggestions for privacy-first households

See it on your own kitchen

A few meal ideas a day, free. No card.