Three-Day Budget Meal Helper：把食材变成三顿晚餐

Prepared for the Hacktoberfest Weekend Challenge: Build for a Friend.
https://dev.to/challenges/hacktoberfest-weekend-2026-10-01

What I Built

A family member wants to keep food spending down but doesn't always know what to cook. Three-Day Budget Meal Helper starts with that everyday problem.

Enter pantry ingredients, a shopping budget, the number of people, foods to exclude, and available cooking equipment. The app puts together three different dinners, one combined shopping list, and a preparation order.

The three-day scope keeps the result concrete. You can see which ingredients are used, what still needs buying, and how the purchases connect to the meals. Shared ingredients can serve more than one dinner without turning all three dinners into the same dish.

The family member has not tried it yet, and no family feedback has been received. This is a working prototype with a real intended user, rather than a claim of proven savings or adoption.

Demo

https://weiweriuser.github.io/three-night-meal-planner/

The app needs an initial download of approximately 136 MB of model assets, plus runtime files. Please allow time for loading, especially on a slower connection or device.

Prices are editable demonstration values, with fixed package sizes. The displayed total is an estimate based on those inputs. It does not represent live shop prices or measured savings.

Code

https://github.com/weiweriuser/three-night-meal-planner

The application code is available under the MIT license. The tested version is commit 6acc565ad3715bbab9cd6386b94f593d35dd80b0.

How I Built It

The AI component matches ingredient phrases to the planner's food vocabulary. It uses paraphrase-multilingual-MiniLM-L12-v2, an open-weight embedding model, running in the browser.

Deterministic normalization and negation handling sit alongside that matching step. The planner then handles dinner selection, ingredient quantities, package rounding, and budget calculations. Recipes and totals come from defined data and algorithms.

That division makes the result easier to inspect. A questionable ingredient match can be examined separately from a shopping calculation. If no valid combination fits the constraints, the app reports that limitation rather than presenting an infeasible plan as successful.

The model has successfully run in the browser. The automated suite passed all 22 tests in two environments. Live browser checks also verified the sample plan, recalculation, cancellation, stale-result clearing, incompatible-unit rejection, and invalid-input handling. These checks do not guarantee correct interpretation of every phrase or compatibility with every device.

Food matching can be wrong. Users should check the recognized ingredients and every recipe. Exclusions are not an allergy-safety guarantee, and the app does not provide medical or nutritional advice.

Why Does Open Innovation Matter?

Open weights make the ingredient-matching layer inspectable and replaceable. The upstream MiniLM model and Transformers.js use Apache-2.0 licensing; ONNX Runtime uses MIT. The repository's third-party notices identify the conversion checkpoint, fixed revision, and license sources.

https://huggingface.co/sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2
https://github.com/huggingface/transformers.js/blob/main/LICENSE
https://github.com/microsoft/onnxruntime/blob/main/LICENSE

Inference happens in the browser, with no application backend or user account. This avoids a hosted inference service for each ingredient query. Hosting and download providers still receive ordinary connection metadata when serving the page and assets, so local inference does not mean zero network contact.

The tradeoff is a substantial initial download and device-dependent performance. Those costs matter for a tool intended to make an everyday task easier.

My Agent Session

Development disclosure: Fully Autonomous AI. I supplied the real-world need; AI wrote and tested the application and prepared this article. Private conversations and identifying family details are not included.

Prize Categories

This project is intended for consideration for the overall prize. No partner-technology category is claimed.