DEV Community: Aman Kumar

"Lazy Downloads: A Sloth-Speed Local AI That Cleans My Messy Folders While I Nap" published: false

Aman Kumar — Tue, 20 Oct 2026 13:25:03 +0000

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

What I Built

Lazy Downloads is a small command-line tool that sorts, renames, and archives the clutter in my Downloads folder. Every file I download ends up named Screenshot 2026-10-12 at 4.31.09 PM (3).png or final_FINAL_v2.pdf, and I never get around to cleaning it up.

It reads each new file, works out what it actually is (an invoice, a lecture note, a boarding pass, a meme), and then:

renames it to something human, like 2026-10-electricity-bill.pdf
moves it into a sensible folder such as Finance/Bills/ or College/Notes/
logs every move to a file so I can undo anything with one command

Like a sloth, it works slowly in the background: it runs as a nightly job, goes through files one at a time, and never hurries my laptop. Everyone with a chaotic Downloads folder gets to be a little lazier.

Demo

Code

How I Built It

Open pieces:

Gemma (open-weight, Google) running locally through Ollama, with a vision-capable variant for screenshots and scanned documents
Python with watchdog for file events, plus small text extractors (PDF text, plain text, image captions)
A plain SQLite file as the move log

Why Does Open Innovation Matter?

My Downloads folder holds bills, ID scans, bank statements, and college files. Uploading those to a closed cloud API just to rename them felt wrong, and an open-weight model made the question go away:

Privacy: every file is read and classified on my own machine, and nothing leaves it.
Offline: it works with no internet, so a nightly job doesn't depend on an API being up.
Cost: after the one-time model download, each run costs nothing, which suits a job that runs every night forever.
Control: I can swap models, edit the prompt, or fine-tune on my own folder taxonomy without waiting for a vendor to change anything.

My Agent Session

The Lazy Payoff

- Files in my Downloads folder before: [N]

Files correctly sorted automatically: [N], sent to _Review/: [N]
Time to run on [N] files on [your laptop specs]: [X minutes]
Time I would have spent sorting by hand: roughly [X minutes]

Prize Categories

- Best Use of Gemma: local Gemma through Ollama is the core of the tool

Best Use of a .tech Domain: only if you claim a free one and host a project page there [keep or remove]
Best Use of GitHub Copilot: only if you actually used it to build this [keep or remove]

NIL

ReceiptDesk: a receipt sorter for freelancers that never uploads a single receipt

Aman Kumar — Tue, 13 Oct 2026 16:52:57 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

What I Built

Every quarter, freelancers dig through a folder of blurry receipt photos and PDF invoices to work out what they spent and what they can deduct. It's tedious, and the paperwork is full of client names, amounts, and addresses they'd rather not hand to a cloud service.

ReceiptDesk is a local app. You drop in a folder of receipts and invoices, and it:

extracts vendor, date, total, tax, and currency
assigns a category (software, travel, equipment, meals, etc.)
flags duplicates and receipts with missing fields
exports a clean CSV for your accountant

It's built for solo freelancers and tiny studios. I think they'd pay a small one-time fee or a cheap yearly licence, because the alternatives are either manual work or uploading client data to a SaaS.

Demo

Code

How I Built It

- Model: Gemma, an open-weight model, running locally through Ollama. [Name the exact size and variant you used.]

Pipeline: [OCR step / vision input, whichever you chose] → the model returns structured JSON → validation → categorisation → CSV export.
Validation layer: the model's output is checked against a schema. Totals must parse as numbers, dates must be real, and anything that fails is flagged for human review instead of guessed.
Stack: [Language, UI framework, storage.]
What I tried that didn't work: [One honest failure, such as a model size that was too slow or a prompt that hallucinated totals, and how you fixed it.]

Why Does Open Innovation Matter?

- Privacy is the product. Receipts and invoices expose clients, income, and locations. With an open-weight model on the user's own machine, I can truthfully say nothing leaves the laptop. That claim is hard to make with a closed API.

The economics work. There's no per-document API bill, so a cheap one-time price still leaves margin. A freelancer with 2,000 receipts a year costs me nothing to serve.
No lock-in. If a better open model appears, I swap one config line. I'm not exposed to someone else's price change or deprecation.
Room to specialise. [If you fine-tuned or prompt-tuned for a niche, say how it did against a general baseline. Only include numbers you measured.]

My Agent Session

My First User

Prize Categories

- Best Use of Gemma

Best Use of a .tech Domain (only if you deployed a landing page on a claimed .tech domain)

N/A

Kindred | Make a difference, close to house.

Aman Kumar — Wed, 07 Oct 2026 13:04:23 +0000

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

What I Built

Demo

Code

How I Built It

Why Does Open Innovation Matter?

My Agent Session

Prize Categories

I Built a Garden Planner with Local Gemma That Ends by Telling You to Go Outside

Aman Kumar — Wed, 07 Oct 2026 12:57:25 +0000

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

What I Built

Frostline is a garden planner that runs entirely on my laptop. You give it two dates, your last spring frost and your first fall frost, and it works out which crops are safe to sow this week and which windows open in the next two weeks. A local Gemma model then turns that into a short weekend plan: a few garden tasks, then one 20-minute job to do outside.

The last line of every plan tells you to close the laptop. Most planners keep you scrolling. This one is meant to end its own session. It's for anyone who wants to start a vegetable patch, or keep one going, without hunting through frost-date charts and forum threads.

Demo

Code

How I Built It

The project has two parts with separate jobs:

Plain Python does the calendar math. Each crop has a maturity time, a frost tolerance, and a sowing lead time. From those and your frost dates, the script computes spring and fall sowing windows and checks them against today's date.
A local open-weight model does the writing. Gemma (gemma3:4b, served by Ollama) gets only the verified facts and is told not to add crops or dates. Its job is to phrase a friendly weekend plan, not to decide what's safe to plant.
I split it this way on purpose. A small local model can be confidently wrong about dates, and a wrong sowing date means a lost crop. Keeping the dates in deterministic code means the worst the model can do is write awkwardly. If Ollama isn't running, the script still prints the facts, so the planner works even when the AI part doesn't.

It uses only the Python standard library, so there's nothing to install besides Ollama.

Why Does Open Innovation Matter?

A garden is often somewhere with weak signal and no reason to ping a server. Because the model runs locally, Frostline needs no connection once the model is downloaded, and no key or account. Where you live and what you grow stay on your machine.

Open weights also let me change the model with one flag (--model) and see how the plan changes. I could add a bigger model for a nicer plan or a smaller one for an old laptop. With a closed API I'd be tied to someone else's pricing and availability.

There are real limits too. A 4B model is a modest writer and the plans can be wordy. The crop table is small and generic, so check your local extension service before trusting a date with a real harvest.

My Agent Session

Prize Categories

Best Use of Gemma

NIL

Family Recipe Keeper

Aman Kumar — Sun, 04 Oct 2026 14:31:44 +0000

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

What I Built

I built Family Recipe Keeper, a local-first app for my grandmother, who keeps recipes in handwritten notes and voice memos scattered across notebooks, phone recordings, and memory. The app helps turn those messy recordings and notes into searchable, organized recipe cards that the whole family can use.

The app:

records and transcribes voice memos using an open-source speech model
extracts recipe ingredients, instructions, and notes from messy text
organizes recipes by meal type, season, or family occasion
keeps everything on the user’s own device instead of sending private family data to a cloud service

It’s a small project, but it solves a real problem for someone I love and makes family traditions easier to preserve.

Demo

[https://your-demo-link.com]

Code

[https://github.com/yourusername/family-recipe-keeper]

How I Built It

I built this around open-source AI tools so it would work locally and keep family data private.

Tech stack:

Open-source local model runtime: Ollama
Speech-to-text: Whisper-style open model
LLM for cleaning and structuring recipe text: Gemma / Llama / Mistral (depending on the variant)
App framework: Next.js or FastAPI
Storage: SQLite
UI: simple local web app

The core idea was to keep the AI entirely local. The app runs on a laptop or home machine without depending on a hosted API, which makes it faster, cheaper, and much more private for a family recipe project.

Why Does Open Innovation Matter?

Open innovation mattered for this project because it made a real difference in privacy, affordability, and control.

A closed API would have required sending voice recordings and personal family notes to a remote server. That isn’t ideal for recipe memories, traditions, and sensitive personal information. By using open-weight models and local inference, I could:

keep the data on-device
avoid recurring cloud costs
swap or upgrade models as needed
customize how the AI interprets informal family notes
build something that works even without internet access

This project is a perfect example of why open-source AI matters: it makes powerful tools accessible, private, and practical for everyday problems that matter to real people.

My Agent Session

Prize Categories

I’m entering:

Best Use of GitHub Copilot
Best Use of Gemma

Not applicable.

Intro

Aman Kumar — Thu, 01 Oct 2026 12:05:31 +0000

Hyy, I am new to the community, happy to be the part of dev.