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

What I Built

Pawelle is a small app that helps my friend Shontu look after her cat, Pinky. Each day Shontu taps through a quick check-in (mood, appetite, energy, play, litter box), and Pawelle writes a short, friendly plan for food and play. It also keeps a diary-style history, so she can see how Pinky has been lately.

The plan is written by a small open-weight AI running on her own computer. There's no account and no cloud, and Pinky's details and photos stay on the machine.

I wanted something simple she would actually use every day. After trying it, this is what Shontu told me:

She loved the app and how easy it is to use. She can now keep track of Pinky in a diary-like way, and the AI meal plan is giving her much better ideas on how to look after Pinky. She's going to use it every day.

Demo

Code

It's MIT licensed. To run it, install Ollama, run ollama pull gemma3:1b, then npm install and npm run dev. The README has the details.

How I Built It

The front end is React with Vite and Tailwind. The back end is Express with SQLite. The AI is Gemma 3 1B, an open-weight model that runs locally through Ollama.

I started with the model's weak spot. When I asked it for an unusual pasta recipe, it invented a dish and fake links, with total confidence. A pet app can't do that, so I kept the model on a short leash:

• Safety rules run first, in plain code. Two days of "not eating", or a danger word like chocolate or lilies in a note, skips the model entirely and shows a calm "please contact your vet" banner.
• The model has a tiny job. It writes one friendly sentence and picks meal times and play ideas from fixed lists. The app writes the rest.
• Everything it returns is checked. Allergens, medicine words, made-up numbers and links are rejected. After one retry, you get a clearly labelled basic plan.
• Free-text notes never reach the model, so they can't be used to steer it.

In a 24-run test with the real model, all 24 plans were valid, with a median time of about 2.6 seconds.

I built it spec-first (a short spec, plan and task list for each feature, all in the repo), with an AI coding assistant doing much of the typing.

Why Does Open Innovation Matter?

• Privacy. A pet's health details and photos never leave the laptop. There's no server to trust.
• No keys, no bills. Everything talks to localhost, with no API keys and no cost per request.
• Swap the model with one setting. OLLAMA_MODEL changes it. I picked a tiny model so it runs on any laptop, and a bigger one is a one-line change.
• I could see and fix what went wrong. The prompt, safety rules and allowed options are plain files in the repo, so when the model misbehaved, I tightened them.

A closed API would have been quicker to set up, but it would have sent a friend's data away and charged per call. Using a small model also made me build the guardrails, and I think the app is better for it.

One honest note: Gemma is open-weight under Google's own terms rather than a standard open-source licence. The app itself is MIT.

Prize Categories

Best Use of Gemma

Pawelle runs Gemma 3 1B, Google's open-weight model, locally through Ollama. It's the only AI in the app. Gemma writes the friendly summary of each plan and picks meal times and play ideas, inside a safety harness: red-flag checks run before it, its choices come from fixed lists, and everything it writes is checked afterwards.

Gemma 3 1B is small enough to run on any laptop, and a bigger Gemma (like 4B) is a one-setting change (OLLAMA_MODEL). The Ollama client and the prompt are in ollama.js and prompts.js.