DEV Community: Aditya Singh Yadav

StoveSide: cook from the next step, not the screen

Aditya Singh Yadav — Sat, 24 Oct 2026 20:40:34 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady.

What I Built

When a recipe is long, cooking can turn into a cycle of touching the phone, scrolling, and finding the line you were on. StoveSide is a small hands-free recipe guide: paste a recipe you already trust, review Gemma’s short steps with source quotes, then have your browser read the current step aloud while you cook.

The aim is simple: make following a recipe feel a little calmer and keep the next step easy to find. StoveSide can also surface questions about ambiguous instructions. A timer appears when Gemma returns a duration; check it against the original recipe before using it.

Demo

[https://drive.google.com/drive/folders/1vZmR8WIu4WR2xkfzcchvyYpjZRZQePoA?usp=sharing]

Code

StoveSide on GitHub

How I Built It

StoveSide uses a dependency-free Node.js server, plain HTML, CSS and JavaScript, browser speech synthesis, and hosted Gemma inference. The server asks Gemma to turn pasted recipe text into short steps with exact source excerpts. It checks that each excerpt appears in the input before returning the draft to the browser. The cook can review each step, listen to it, and optionally start its timer.

The test suite has three tests for source-quote validation, timer bounds, and invalid or empty results. These are software checks, not a measurement of model accuracy. The app is configured to use gemma-4-26b-a4b-it through Google’s hosted API. Verify a real request with a non-sensitive recipe and confirm the browser flow before publishing.

Why Does Open Innovation Matter?

Gemma’s open weights give builders the freedom to adapt the model or choose a different way to run it. StoveSide currently uses hosted inference to keep setup lightweight; when someone requests steps, recipe text is sent to Google’s API. The app keeps the API key on its local server and does not write submitted recipe text to disk. A future local-inference option could keep recipes on the cook’s device, though I have not built or evaluated that mode yet.

StoveSide is a prototype, not a substitute for the original recipe or food-safety guidance. Generated steps can be wrong or incomplete, so review them against the source before cooking.

Footpath Receipt — Small observations, clearly recorded

Aditya Singh Yadav — Sat, 10 Oct 2026 08:51:07 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass.

What I Built

Footpath Receipt turns details noticed on a familiar walk into clear, reusable field notes. I wanted a lightweight way to capture what I observed and organize it later, while keeping the person who took the walk in control of the result.

Write down what you personally noticed, and Gemma organizes those notes into an editable field receipt. Each observation points back to a note ID and an exact quote, with follow-up questions kept visible. You can edit, copy, or download the result.

The goal is to make the walk the important part and keep the screen task small: capture a few notes, organize them, then review the draft.

Demo

[https://drive.google.com/drive/folders/1vZmR8WIu4WR2xkfzcchvyYpjZRZQePoA?usp=sharing]

Code

Footpath Receipt on GitHub

How I Built It

The app uses plain HTML, CSS, browser JavaScript, and Node.js’s built-in HTTP server, so it starts without a framework or package installation. The server sends the notes and optional manually typed area label to Google’s hosted Gemini API, using the open-weight Gemma 4 26B A4B Instruct model (gemma-4-26b-a4b-it). The API key stays on the local server and isn’t included in the repository or browser code.

To keep results grounded in the author’s notes, the server checks each note ID and verifies that its quoted source text appears in the cited note. The interface shows that quote beside each organized observation, making it easy to review and refine the result.

The automated test suite passes four tests covering note validation, citation mapping, rejection of invented sources, and malformed model output. I also ran a real Gemma request on my Windows laptop and received a response, confirming the hosted inference flow end to end.

Why Does Open Innovation Matter?

Gemma’s open weights give developers the option to adapt or self-host the model under its terms. I chose hosted Gemma for this version to keep setup simple—there’s no model download. In hosted mode, the app sends the notes and optional area label to Google for generation, while keeping the API key on the local server. This gives the project a quick way to try open-weight AI today and a path to local deployment or adaptation later. The author reviews the cited draft before deciding how to use it.

Prize Categories

Best Use of Gemma — Gemma powers the project’s central task: organizing first-hand notes into a cited, editable field receipt. I verified the hosted inference flow with a real request on my Windows laptop.

Pakka? - Plans, with receipts

Aditya Singh Yadav — Sun, 04 Oct 2026 21:19:52 +0000

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

What I Built

A friend recognized the problem behind Pakka?: important plans get buried in busy group chats, tentative suggestions can look final, and later changes are easy to miss. I built a prototype to turn pasted chat text into reviewable plan candidates.

I shared Pakka? with my friend, who already had Ollama installed. He ran it locally on his device, got it working, and genuinely liked the project. That was encouraging: it showed me this idea mattered to someone beyond me.

Pakka? separates confirmed, proposed, unclear, cancelled, and superseded plans. It shows the source lines behind each interpretation, keeps tasks and RSVPs out of calendar events, and requires a person to review an event before export.

Demo

Drive folder

Code

View Pakka? on GitHub.

How I Built It

Pakka? is a Node.js app with a Gemma 4 integration through Ollama. It asks the model for plan candidates with source quotations and line numbers, then checks that cited text appears in the supplied conversation. A person can edit and review candidates before exporting valid, confirmed calendar events. My friend ran it locally using his Ollama setup, and it worked on his device.

The automated test suite passes 13 tests covering evidence checks, dates, cancellations, review requirements, and calendar export. This is a small software test suite, not a broad measure of model accuracy.

Why Does Open Innovation Matter?

Ollama let my friend run Pakka? on his own device instead of having to use a hosted model service. In local mode, chat text goes to the local Ollama instance; the optional hosted mode sends it to Google's Gemini API. That gives people a choice about where inference happens. The local run worked for my friend, though I have not measured model accuracy or compared it with a closed model.