Pantrymind
PantryVision-3 + Chef-LM 2 · fridge-to-recipe AI

Your fridge, seen by AI. Tonight's dinner, solved.

One photo or a typed list. A vision model detects every ingredient, a constraint solver applies your diet, time and servings, and a recipe model writes three meals that use up what you already own.

Ingredient classes
1,800+
Detect latency p50
182 ms
Photo to 3 recipes
~7 s
Vision photo or text inputSolver diet · time · servingsOutput 3 recipes + macrosPrivacy photos discarded after detection
Live demo · kitchen lab

Run the model on your kitchen.

We've stocked a sample fridge. Edit any item, set your constraints and watch every stage of the pipeline run — from detection to macros.

Connecting to live AI…
pantry-01.session PantryVision-3 → Chef-LM 2 → NutriNet 0/5 runs
Agent traceidle
    Model streamidle
    tokens 0tok/s 0elapsed 0.0 s
    // Raw JSON from Chef-LM 2 streams here when you run the pipeline.
    Output · awaiting run

    Three recipes, an ingredient graph and a nutrition estimate.

    Every recipe shows which of your items it uses, what you'd need to buy, macros per serving, how much perishable food it rescues and whether it passes your diet, time and servings constraints.

    How it works

    Fridge to fork in three moves.

    01 · SENSE

    Show us the shelves

    Take one photo of your fridge or pantry, or type a quick list. Our vision model labels items, even half an onion in a bag.

    02 · CONSTRAIN

    Set the house rules

    Servings, time you've got, diet and the cuisine you're craving. Pantrymind remembers allergies and dislikes per profile.

    03 · GENERATE

    Cook, tick, done

    Pick a recipe ranked by how much it uses up. Follow steps you can check off, and get a tiny shopping list only when it's worth it.

    Photo pantry scan

    Recognises 1,800+ ingredients, including jars, leftovers in tubs and regional produce like garden egg and ugu.

    mAP@.5 0.89

    Use-it-up ranking

    Recipes are scored on how much of your stock they use, prioritising things close to their best-before.

    waste-weighted scoring

    Minimal shopping lists

    Only suggests buying something if it unlocks a much better meal — and merges every gap into one list.

    median 0.6 items to buy

    Diet you can trust

    Halal, vegan, vegetarian and low-carb rules are checked against every ingredient, reviewed by our registered dietitian.

    ingredient-level verifier

    Honest timings

    Times include chopping and preheating. Ask for 15 minutes and you get dishes that really fit 15 minutes.

    hard time constraint

    Weekly meal plans

    Plus and Family plans stitch recipes into a week that shares ingredients, so one bunch of spinach covers three meals.

    multi-day planner
    Under the hood

    Three models, one knowledge graph, hard guardrails.

    Detection, planning and nutrition are separate specialised models. Every recipe passes a rule-based verifier before you see it, so the language model never has the last word on your diet.

    INPUTPERCEPTIONPLANNINGGUARDRAILSOUTPUT Fridge photoJPEG/HEIC, 1 frame Typed listfree text, any order Household profilediet, allergies, time PantryVision-3detector · 1,800 classes Ingredient normaliserontology + unit parser Constraint solverdiet · time · servings Chef-LM 2recipe generator, 64k ctxNutriNet macro headwaste-weighted ranking Knowledge baserecipe graph · 240k dishesnutrition DB · 9.2k foodsdiet rules · RD-reviewed Verifierallergen checkdiet rule matchtime sanitymacro boundsunsafe tempsreject + regen 3 recipesmacrosshopping listwaste score
    Ingredient recall (photos)
    93.4%PantryVision-3 on 12k held-out fridge photos
    Latency p50
    182 msdetection; full plan ~7 s end to end
    Diet rule precision
    99.7%verifier vs dietitian-labelled set
    Context window
    64ktokens: pantry, profile and retrieved recipes
    Languages
    11incl. Yoruba, Igbo, Hausa ingredient names
    Privacy
    0 daysphoto retention. No training on your data.

    Internal benchmarks, Pantrymind Labs, September 2026. Fictional figures for this demo site.

    From real kitchens (well, fictional ones)

    Less binned. More dinner.

    38%less food thrown away by weekly users after 8 weeks
    4.1average meals rescued per household, per week
    $27saved on groceries per month, median
    7sfrom photo to three recipes

    "Typed in half a cabbage, rice and two sad sausages. Got a fried rice I'd actually make again."

    Tola K. · nurse, Lagos (fictional)

    "The halal filter is the first one that's never slipped something past me. Ramadan evenings got easier."

    Yusuf R. · dad of three, Leicester (fictional)

    "My shopping list went from a page to four lines. Plantain doesn't go black on my counter anymore."

    Gabriela M. · grad student, Austin (fictional)
    Pricing

    Priced like a bag of groceries.

    Free

    For trying it on tonight's dinner.

    $0/mo
    Free forever
    • 5 recipe runs a week
    • Typed ingredient lists
    • All diet filters
    • Shopping list copy
    Most kitchens

    Plus

    For cooks who want the fridge to do the thinking.

    $6/mo
    Billed monthly
    • Unlimited recipe runs
    • Photo fridge & pantry scan
    • Weekly meal plans
    • Saved recipes & best-before nudges

    Family

    For one fridge, many opinions.

    $14/mo
    Billed monthly
    • Everything in Plus
    • 6 profiles with diets & allergies
    • Weekly family plan
    • Shared, live shopping list

    Paid plans start with a 14-day trial. No card required — we'll ask before anything is charged.

    The team

    Chefs, scientists, one very picky dietitian.

    FA

    Funmi Adeyemi

    CEO & co-founder

    Ran a Lagos supper club for six years, then spent three researching household food waste.

    @funmicooks · in/funmi-adeyemi
    MB

    Marco Bellini

    CTO & co-founder

    Computer-vision engineer who taught a model to tell spring onions from leeks in bad fridge light.

    @marcosees · gh/mbellini
    AN

    Amara Nwosu

    Nutrition lead, RD

    Registered dietitian who signs off every diet rule and calorie model before it ships.

    @amara.eats.well · in/amaranwosu
    DW

    Dan Whitfield

    Product designer

    Designs for greasy thumbs and propped-up phones. Previously built checkout flows for grocers.

    @danwhit · dribbble/danw
    FAQ

    Questions from the queue.

    Something else? Write to hello@pantrymind.app and a human replies within a day.

    How accurate is the photo scan?

    On clear fridge photos it labels around 9 in 10 visible items correctly. You always see the detected list and can edit it before recipes are made, so a mislabelled jar never ends up in dinner.

    Can I trust the diet filters for halal or vegan?

    Every recipe is checked against ingredient-level rules written with our dietitian — gelatine, stock cubes, wine and fish sauce included. If an item is ambiguous, Pantrymind flags it instead of guessing.

    Do I need special ingredients?

    No. We assume salt, oil, water and black pepper. Anything else a recipe needs shows up on the shopping list, and we rank recipes that need nothing extra first.

    Are the calorie and macro numbers exact?

    They're estimates per serving based on typical portion weights. Good for planning, not for medical dosing. Plus users can log real quantities for tighter numbers.

    What happens to my fridge photos?

    Photos are used to detect ingredients and then discarded. We keep the ingredient list, not the picture, unless you choose to save it. We never train on your data.

    Can I cancel anytime?

    Yes, from settings, in two taps. Annual plans are refunded pro-rata within the first 30 days.

    Tonight's dinner is already in your fridge.

    Join the list for the mobile app and early Family plan access.