DEV Community: Shreyash Magar

Midnight growls!!

Shreyash Magar — Sat, 31 Oct 2026 06:05:57 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine

What I Built

Abuela’s Recipe Cards is a candlelit ofrenda for one family recipe: Strawberry Pumpkin Patch Brownie Cake.

It’s a single, offline-first web page set in a pumpkin patch at night. Pumpkins glow like little lanterns along the bottom, strawberry vines creep up the edges where cobwebs would normally go, and marigold petals drift down the screen. Ingredients are laid out as offerings on the ofrenda, and the steps are a path through the patch. The recipe’s original wording is kept exactly as it was written.

The part that makes it more than a theme is the ghost guesses. Family recipes are full of instructions like “a pinch of cinnamon,” “a splash of milk,” and “bake until it looks right.” Those measurements lived in someone’s hands, not on the card. On my page, every vague phrase glows faintly. Hover over it and a local open-weight model whispers a suggested measurement, clearly labeled as a guess, so the original is never overwritten.

I read the theme two ways. The spooky side is the pumpkin patch, the ghost ink, and the whispering. The tender side is the point: Día de los Muertos says a person lives on as long as someone remembers them, and a recipe is one of the most practical ways to keep remembering. [One line: who this recipe belongs to and why it matters to you.]

How I Built It

Model: [Qwen2.5 7B / Gemma / Llama, your actual choice], running locally through [Ollama / llama.cpp]. No API keys, no network calls.
Step 1, find the vague parts. The recipe text goes to the model, which flags imprecise amounts and timings (“a splash,” “until the top cracks,” “enough strawberries”) and returns them as structured JSON.
Step 2, suggest, never overwrite. For each flagged phrase, the model proposes a concrete measurement, a confidence level, and a short reason. These are stored beside the original wording, never in place of it.
Step 3, render the ofrenda. The JSON is baked into one self-contained HTML file. The candle flicker, drifting marigolds, strawberry vines, and glowing pumpkins are all CSS and a little vanilla JS. There is no framework, no CDN, and no tracking.
Optional extension: the same pipeline can take a photographed handwritten card through a vision-language model such as Qwen-VL and feed it into the same format. [Keep only if you built or tested it.]

The result is one HTML file that opens with the internet off.

Why Does Open Innovation Matter?

It’s private by default. A family recipe comes with a name, a story, and handwriting that can’t be replaced. Running the model locally means none of that is uploaded to a service I don’t control.

It works with the lights off. The model runs on my own machine, and the finished page is a single file. That suits a project about remembering.

It can’t be taken away. Hosted services change pricing, retire models, and shut down. Open weights stay on disk, and plain HTML will still open in any browser decades from now. A recipe meant to outlive a person shouldn’t depend on an API that might not outlive the year.

I control the behavior. With a local model I can decide how cautious the ghost is: whether it guesses boldly or says “I don’t know,” and how it labels uncertainty. For a family heirloom, an honest “this is a guess” matters more than a confident wrong answer, and I could tune that myself.

SlothSort: My Idea for a Lazy, AI-Powered File Organizer

Shreyash Magar — Sun, 25 Oct 2026 12:44:13 +0000

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

What I Plan to Build

My computer's Downloads folder can quickly become a mess of PDFs, screenshots, documents, installers, images, and files with confusing names. Finding a particular file often takes more time than it should.

That is why I want to build SlothSort — a local AI-powered desktop assistant that organizes files so you don't have to do it manually.

The idea is to let users scan a cluttered folder and receive suggestions for how to organize it. SlothSort would help with:

Smart categorization: Group files into categories such as documents, images, videos, code, and installers.

Suggested file names: Propose clearer names that make files easier to recognize.

Automatic folder planning: Recommend folders and locations based on file types and names.

Approval before changes: Show users a preview of the proposed changes and let them approve or reject them before anything is moved or renamed.

The goal is to reduce digital clutter without making users spend their time sorting hundreds of files themselves.

SlothSort is currently a project proposal. I plan to develop a working prototype and gradually add more intelligent organization features.

Code

The code repository is not available yet. I plan to publish the project on GitHub as development progresses.

How I Plan to Build It

My plan is to combine Python file automation with a locally running open-weight language model through Ollama.

The proposed workflow is:

The user selects a folder to organize.

Python scans the folder and collects basic file information, such as file names, extensions, and sizes.

The system groups recognizable file types using simple rules and asks the local AI model to help classify ambiguous files or suggest more descriptive names when useful.

SlothSort presents a preview of the proposed folder structure, file names, and changes.

The user reviews and approves the changes before Python moves or renames anything.

I want to build the first version around a simple principle: the AI suggests, but the user stays in control.

The initial version will focus on file metadata rather than reading the contents of every file. Any additional content analysis can be introduced later with clear user consent.

Once the required software and model are installed, I want the core workflow to run locally without requiring a continuous internet connection.

Why Does Open Innovation Matter?

File organization might seem like a small problem, but a folder can contain personal documents, screenshots, work files, and other information that users may not want to upload to an external service.

That is why I want local, open-weight AI to be central to SlothSort.

Privacy: File names and metadata can be processed locally, avoiding the need to send them to a third-party AI API.

Offline capability: Once installed, the core organization workflow could work without a continuous internet connection.

Flexibility: I can experiment with different open-weight models and choose one that performs well on the available hardware.

Lower ongoing costs: Running inference locally avoids per-request AI API charges, although the user still needs suitable hardware and must account for electricity and storage.

A closed AI API could provide similar classification features, but it could also introduce external-service dependencies and usage costs. With SlothSort, I want users to be able to run the assistant on their own machines and decide exactly which file changes to accept.

Open innovation would also let me learn from the implementation, modify the workflow, and adapt the tool to different users' needs.

My Agent Session

There is no agent session yet because development has not started.

As I build the project, I plan to document the development process and share a session or walkthrough if it is useful.

The Lazy Payoff

The payoff I am aiming for is simple: a cleaner computer with less repetitive work.

Instead of manually creating folders, renaming files, and moving downloads one by one, users could review a proposed organization plan and approve the changes in a few steps.

I have not measured any time savings yet because SlothSort is still at the idea stage. After building the prototype, I want to test it on a sample folder and compare the time required for manual organization with the time required using SlothSort.

Safety is an important part of the idea, too. The tool should preview changes, avoid deleting files automatically, and give users control over when changes happen.

What's Next?

My first milestone is a small, functional prototype that can scan a selected test folder, suggest categories, and safely organize files after user approval.

From there, I hope to introduce smarter file-name suggestions, customizable categories, and more advanced organization rules.

This is my idea for Slow and Steady: a small, practical use of open-source AI that handles a tedious task while leaving the final decisions in the user's hands.

Golden Hours

Shreyash Magar — Sun, 11 Oct 2026 06:29:05 +0000

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

A team project by @shikha_singh_20 and @shreyash_magar.

What I Built

Golden Hour is a calm, mobile-first web app that nudges you to step outside around sunrise and sunset.

You pick sunrise, sunset or both, and how many minutes ahead you want a reminder. When it's time, your phone gets a notification such as "Golden hour in 20 minutes. Take your chai out to the terrace." You go to your balcony or terrace, take one photo of the sky, and enjoy a chai, coffee or snack. Then you tap Done. The check-in is recorded, your streak goes up, and you put the phone away.

It's for people who spend long hours on screens and for people in apartments, where "going outside" can simply mean going to the balcony. The screen is the shortest part of the experience: a short setup, and a tap on Done.

Code

[https://github.com/shikha-code404/golden-hour]

How I Built It

Open-source pieces

Gemma (open-weight), run locally through Ollama. We used it to write the reminder messages, treat suggestions and completion lines. We generate them with scripts/make_nudges.js, review them by hand, and save them to data/nudges.json. The server picks lines from that file, so the live app doesn't need a model running.
SunCalc calculates sunrise, sunset and golden hour offline from latitude, longitude and date.
Node.js and Express for the server, and Web Push with VAPID keys for notifications. The frontend is a plain-JavaScript PWA with a service worker.
MongoDB Atlas stores subscribers, reminders and check-ins.

How it works

You allow notifications and share your location. The app stores your timezone too.
A scheduler calls our /api/tick endpoint every minute with a secret header. The endpoint plans upcoming reminders from SunCalc and sends the ones that are due.
Tapping the notification opens an "outside" screen with the camera, a treat picker and a Done button.
Your photo is saved in your phone's IndexedDB and is never uploaded. The server only stores the check-in (date, kind and optional treat) to count your streak.

Data model: subscribers (push endpoint and keys, location, timezone, preferences, lead time), reminders (kind, event time, fire time, sent time) and checkins (kind, local date, optional treat).

Hosting: Render runs the web service, defined in a render.yaml Blueprint with a /health check and secrets entered in the dashboard. Render is only hosting. Any server with Node.js and MongoDB can run the app.

How we built it: We used Antigravity as a coding assistant and built the app in phases (database, sun-time module, API, frontend, messages, deployment), checking each step ourselves.

Why Does Open Innovation Matter?

We could read and curate every message. Because Gemma is open-weight and ran on our own machine, we generated the messages, deleted the odd ones, and shipped a fixed set. We aren't relying on a hosted model whose wording or terms could change.
No per-use AI cost. The model runs once, offline, on a laptop. There are no third-party AI API keys, so the AI part of the app costs nothing to run. (Web Push and MongoDB Atlas still need their normal service credentials.)
Privacy by design. Photos of someone's home never leave their phone, and sun times are calculated with an open library instead of sending the person's location to a sun-times API.
Swappable and extensible. When a better small model appears, we can regenerate the messages and compare. Anyone can fork the repo, edit data/nudges.json, and add their own language or tone.

Prize Categories

We have used Gemma , Render, MongoDB Atlas for building this project.