Fridge First and the small kindness of making dinner

Manish Kumar — Sun, 04 Oct 2026 20:23:59 +0000

This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend.

What I Built

“Have you eaten?”

That is the feeling I wanted Fridge First to have. An ordinary question, with care tucked inside it.

I built Fridge First for someone I love. Cooking for them and sitting down to eat together makes me happy. I didn’t start with a problem to fix. I started with wanting to make a small thing that carries that care into an ordinary evening. Fridge First begins with what’s already in the kitchen, keeps the little notes that make a recipe feel like yours, and makes it easy to say, “Come over, I’ll cook.”

I don’t have feedback from them to share yet, so I won’t put words in their mouth.

Fridge First is a personal kitchen notebook. You add what you have, and Gemma helps suggest three things you could cook. You can keep a recipe, leave a note after making it, and find that note beside the recipe the next time you come back.

The part I care about most happens after the suggestion. “More lemon next time” starts turning a recipe into your recipe. A supper postcard keeps a few words from the evening. An invitation lets you say, “Come over. I’ll cook,” and ask a friend to bring something you’re missing.

The app began with the fridge, but I wanted the whole experience to hold that personal feeling. So “I made this” opens with a place for your note. Updating the ingredient list sits underneath. The recipe book brings back what you wrote, instead of making every dinner feel like starting again.

Demo

The app currently runs locally; a public deployment has not been verified.

A short way to explore it:

Open the app and choose “Have a peek at an example.” The recipes are labelled samples, and no model download starts.
Open a recipe and keep it in your book.
Choose “I made this,” write a note, and save the supper.
Open its postcard, then return to the recipe. Your note comes with it.
Choose “Invite a friend,” select what they could bring, and add your own message.

New suggestions use the “What could I cook?” button. That starts the local Gemma worker and, on the first use, a model download. The sample flow demonstrates the surrounding app; it is not evidence of live model output.

Code

Fridge First on GitHub

The application is MIT licensed. The repository contains the app, local server, tests, setup instructions, and dependency credits. The submission write-up lives separately.

How I Built It

The app is plain HTML, CSS, and JavaScript. It has no build step or package dependencies. I worked with Codex on the implementation and iterated on the product until the personal touches reached the everyday flows.

For recipe generation, the app uses Gemma 3 1B through WebLLM in a browser Web Worker. The selected model is gemma3-1b-it-q4f16_1-MLC. It starts only when someone asks for new ideas, and there is an explicit stop button to release the worker.

The model receives the ingredient list, cooking time, and number of people, and returns structured recipe data. Application code validates the result before showing it. It also checks which ingredients are on the list; the model does not get to decide that something is available.

I kept remaining quantities manual. “Half a bag” is useful to a person, but it is a poor basis for pretending to subtract an exact amount. After cooking, you can leave an ingredient alone, remove it, or change what is left. Undo will not overwrite newer fridge edits.

The notebook stays in browser storage. Postcards are drawn to a canvas for image export, and invitations can be copied with the full recipe. Nothing is sent to a friend automatically.

Why Does Open Innovation Matter

For this app, the useful consequence is that recipe inference can happen on the cook’s device. Their ingredient list does not need to go to a hosted inference API, and they do not need an API key or an account to use the notebook.

Gemma provides accessible model weights under its own terms, and WebLLM provides the open-source browser inference runtime. The application code is available to inspect and change: how ingredients are prioritized, how recipes are validated, and what the interface does with the result.

There are tradeoffs. The first model download takes time, compatible graphics support is required, and this is not a promise that the entire app works offline. A different model would need compatibility work and testing. Those limits matter, especially for something intended to make an ordinary evening easier.

The result I want is small: a little less effort deciding what to cook, and a place to keep the details that make it yours.

Prize Categories

Best Use of Gemma. Gemma 3 1B is the model integrated for local recipe generation through WebLLM. Live inference verification remains outstanding before publishing this draft as a completed submission.