Off My Plate

Online grocery shopping is solved. Deciding what to cook is not.

The problem

Households spend real time and mental energy every week answering "what are we eating?" under constraints that fight each other: allergies, picky eaters, differing tastes, the time actually available to cook, a budget, and whatever is already in the fridge.

Recipes, preferences, pantry contents and the shopping list all live in different places, and none of them talk to each other. Grocery apps begin after that work is done — they help you buy things, not decide.

What we built

Off My Plate — a meal-planning feature that lives inside an existing online grocery app, taking the decision off the customer's plate and turning it straight into a filled basket.

Knows your household → knows what you have → plans what you should eat → lets you adjust it → buys only what's missing.

It runs end to end: a real backend, a real conversational planner, and a real basket hand-off.

The experience

Surface	What it does
Household	A profile per person — allergies, favourites, dislikes, age. Allergies are hard constraints; preferences are strong signals.
Ask	Plain language in, plan out: "dinner tonight, about 30 minutes." Voice input included.
Week plan	All 28 meals of a week as a grid — cook times, who eats each meal, and labels for kid-friendly, vegetarian, and leftovers.
Recipes	Saved from anywhere: the retailer's own, your own, or imported by pasting a link or photographing a page.
Pantry	Fridge, freezer and cupboard with low-stock flags — kept current from past orders, receipts, or just by telling it.

Two ideas do most of the work

One meal, small variations. Rather than cooking separately for everyone, the planner finds a single meal that works for the whole table. Asked to feed a household where one adult likes salmon, another likes steak, and two young children are suspicious of vegetables, it produced:

Lemon-Garlic Salmon (& Steak) with Rice and Plain Broccoli
One method, three plates.

Preferences are learnable, not just editable. Saying "the six-year-old doesn't like mushrooms" in the chat updates that profile immediately. Saying "we've got rice and four eggs" stocks the pantry. It gets more useful without anyone filling in forms.

How it works

• Two planning modes. A fast single-call mode answers in about five seconds. A multi-agent mode has one model orchestrate while separate agents draft, independently verify against the household's constraints, and revise — slower, with a second opinion on every plan.
• Constraints hold under pressure. Asked directly for a peanut dish for a nut-allergic household, it refuses and names who it would affect rather than quietly substituting.
• Pantry-aware basket. An approved plan becomes a basket by subtracting what's at home, matching the rest to the catalogue with real prices, and splitting the result three ways: in the basket, already in your kitchen, and not stocked.
• Catalogue independence. Recipes aren't limited to what the retailer sells. Unavailable items are surfaced honestly at basket time, not hidden by pruning the recipe.
• Hand-off, not replacement. Confirming pushes the basket into the customer's real grocery basket and returns a checkout link. The retailer keeps the transaction.

State of the build

Running as a full stack from a single command. The whole chain is covered by automated browser tests that drive the real interface against the real backend: household saved, plan generated, recipe rendered, basket built with correct pantry subtraction and pricing, quantity changed, basket confirmed, pantry updated.

Built on the host app's own design system — colours, typeface, spacing, component patterns — so it reads as part of the product rather than a bolt-on. Responsive from phone to desktop, light and dark themes, and packaged as a mobile shell as well as a web app.

Why it matters

The feature converts a question customers answer alone — and often badly, at 6pm — into a completed grocery order.

• Bigger baskets, because a week of meals is a longer list than a top-up shop.
• Higher frequency, because planning becomes a weekly habit inside the app.
• Better retention, because a profile that knows your allergies and your children's tastes is costly to recreate elsewhere.
• Less waste, because plans start from what's already in the kitchen.

What we'd do next

• Close the loop with light post-meal feedback, so the system learns which recommendations actually worked.
• Real photography in place of the illustrated ingredient set.
• Receipt and fridge-photo capture, so pantry data needs no manual upkeep.
• Spend, nutrition and waste trends once enough weeks accumulate.
• Treat the household profile as governed customer data — consent, export, deletion — given it holds allergy information.