DEV Community: Deependra Gurjar

SlothDesk: A Lazy, Offline AI That Cleans Your Downloads Folder

Deependra Gurjar — Sun, 25 Oct 2026 09:17:02 +0000

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

What I Built

SlothDesk is a private, local-first AI file organizer. It reads a short text preview of the files in a messy folder (my Downloads folder was the inspiration), decides where each one belongs, and suggests a clear, descriptive filename.

Everyone has a folder full of scan_0043.pdf, final_final_v2.docx, and random invoices. Sorting them by hand is exactly the kind of chore a sloth would avoid. SlothDesk does the thinking for you, so you only have to look over the plan and click apply.

It's also built for people who don't want their documents going anywhere. Invoices, ID scans, and personal papers are the files you least want to upload to a cloud API, so SlothDesk runs entirely on your own machine.

What it does:

Reads the first 600 characters of .pdf, .txt, .md, .csv, .json, and .py files to understand what they actually are
Sorts each file into one of seven categories: Invoices & Receipts, Work & Documents, Code & Dev, Personal & Identity, Media & Assets, Archives & Installers, or Miscellaneous
Suggests a descriptive new filename for each file
Shows a dry run table with every proposed move, rename, and the AI's reasoning, so nothing changes until you approve it
SlothDesk runs locally, so there's no hosted link. You can try it yourself in a few minutes with the steps in the README.

Code

DeependraGurjar / SlothDesk

Private, local-first AI file organizer. Reads PDF and text snippets offline via Ollama to automatically rename and sort your Downloads folder without cloud APIs.

🦥 SlothDesk Desktop Cleaner

Clean your messy folders with 100% offline, open-source AI.

SlothDesk is a local-first file organizer built with Python, Streamlit, and Ollama. It reads a short preview of each file, uses a local LLM to decide where it belongs, and suggests a clear filename. Nothing is uploaded anywhere, and nothing moves until you approve it.

Features

Offline and private: runs on your machine through Ollama. No cloud APIs.
Content-aware: reads the first 600 characters of .pdf, .txt, .md, .csv, .json, and .py files instead of guessing from the extension.
Reliable output: Pydantic schemas keep the model's answers clean and valid.
Dry run first: preview every proposed move, new filename, and the AI's reasoning before anything changes.

Category	Examples
Invoices & Receipts	bills, receipts, payment confirmations
Work & Documents	reports, notes, letters
Code &

How I Built It

The whole project is built around open-source tools that run on a normal laptop.

Open-weight models via Ollama: SlothDesk uses small 3B models, qwen2.5:3b or llama3.2:3b, served locally by Ollama. They're small enough to run without a GPU, which fits the slow-and-steady theme.
Structured outputs with Pydantic: Small models can ramble or return messy text. I defined a Pydantic schema for the response (category, new filename, reasoning) and pass it to ollama.chat as the output format. That keeps the model's answer valid JSON with a category from the fixed list, so the app can trust the result instead of parsing free text.
Content previews with pypdf: For PDFs, pypdf pulls out the first 600 characters of text. Text-based files are read directly. Giving the model a snippet instead of the whole file keeps things fast and keeps the prompt small.
Streamlit UI: A simple Streamlit app lets you pick a folder, run the analysis, and review the dry-run table before applying anything.

The flow is: scan the folder, extract a preview, ask the local model for a structured answer, show the plan, and only then move files.

Why Does Open Innovation Matter?

For this project, open models weren't just a nice-to-have. They're the whole point.

A file organizer needs to read the contents of your files, and those files are often your most sensitive ones: invoices, certificates, identity documents. With a closed API, using AI for this means sending those documents to someone else's servers. With open-weight models and local inference, nothing leaves your machine. It also works offline, has no API bill, and has no rate limits.

Open tooling also made the project possible for a student on a normal laptop. Ollama, small open models, and Pydantic are all free and well documented, so I could build and test the idea without any paid service. And because the repo is open, anyone can add file types, categories, or support for other models.

Team

Built by me and my teammate Richa Kumari (GitHub).

ColdCraft Studio: Personalized Cold Emails, Written Locally for $0

Deependra Gurjar — Sun, 18 Oct 2026 16:03:40 +0000

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

What I Built

Everyone has received the cold email that opens with "I hope this email finds you well." It goes straight to the trash.
ColdCraft Studio fixes that. Give it a company's website URL and a short description of what you offer, and it writes a cold email that actually references what the company does. No template, no 15 minutes of research per lead, and no cloud AI bill.

Here is what happens when you click Generate:

It scrapes the company's site for taglines, current initiatives, and mission statements.
A local open-weight AI model reads that content.
It drafts a short email built around specific details from the page.
Everything runs on your own laptop.

Who is this for?

Freelancers and indie consultants pitching web development, design, content, or technical services to businesses and colleges, without spending 20 minutes researching every lead.
Sales reps and lead-gen agencies who need hundreds of tailored emails a week without paying $50 a month in token bills.
The Real Problem
Cold email only works when it feels personal, but right now there are two bad options:

What happens

Cloud AI looks like the answer, but running 1,000 detailed web pages through GPT-4o costs roughly $30 to $50 a month. You are also uploading private pitches and prospect lists to someone else's servers.
ColdCraft does the research and drafting in under 5 seconds per lead at $0.00 marginal cost.

Key features

Realistic email client preview. The output does not look like a chatbot. It renders inside a simulated desktop email window, with a proper salutation, body paragraphs, and sign-off.
One-click dispatch. One button opens your default mail app (Outlook, Apple Mail, and others) with the recipient, subject, and body already filled in, using a mailto: link.
CSV export. Save leads, subject lines, and personalized bodies to a spreadsheet for your pipeline.
Code

DeependraGurjar / coldcraft-studio

A local-first, zero-cost B2B cold outreach engine powered by Ollama and open-weight models.

ColdCraft Studio

A local-first B2B cold outreach tool. It reads a prospect's website, finds a relevant talking point, and drafts a personalized cold email using open-weight models running on your own machine through Ollama. No API keys, no per-email cost, and no data leaves your computer.

Overview

Effective cold outreach is slow and expensive to do well:

Mass templates read as robotic, often land in spam, and convert poorly.
Manual prospect research takes 10 to 15 minutes per lead.
Cloud AI APIs speed up drafting but add per-request costs and send your pitch and lead data to third-party servers.

ColdCraft Studio automates the research and drafting steps locally. Give it a prospect URL and a short description of your offer, and it produces a ready-to-review email, typically within a few seconds on modern hardware…

View on GitHub

How I Built It

The stack is small and local-first. After the page is scraped, everything runs offline.

URL → Scraper → Studio → Local AI → Email

Prospect URL: You paste in the company's website.
BeautifulSoup scraper: Strips menus, footers, cookie banners, and scripts, keeping only headings and paragraphs.
Streamlit studio: You add your offer, goal, and tone.
Ollama + qwen2.5:3b: Generates the email as strict JSON.
Output: Email preview, mailto: link, and CSV export.

Local AI engine. I used qwen2.5:3b through Ollama, with an option to switch to llama3.2:1b on lighter machines. Qwen 2.5 is the star of the project. Even at 3 billion parameters, its instruction following and JSON formatting are rock solid.

Smart web scraping. The scraper is written in Python with BeautifulSoup. It removes navigation bars, footers, cookie banners, and scripts, and keeps only the high-signal headings and paragraphs. That gives the model something worth writing about.

Deterministic output. I forced Ollama's format="json" and paired it with a strict prompt framework:

Icebreaker, then Concrete Problem, then Value Prop, then Soft CTA

This stops the model from hallucinating or producing one long, unreadable block of text. Every email follows the same proven shape.

Frontend. The interface is built with Streamlit and custom CSS, so it feels like a focused desktop studio instead of a standard AI demo.

Why Open Innovation Matters

This project is a good example of why open-weight models matter for indie developers and micro-businesses.

Unit economics that work. Say you charge a freelancer $15 a month for an outreach tool. If cloud API bills cost you $12 a month per active user, your margin is gone. With qwen2.5:3b running locally, the marginal server cost is $0.00.

Real privacy. Pitches often involve unreleased projects, internal proposals, and private contact lists. With an open model, none of that leaves the user's hard drive.

No moving targets. No rate limits, no surprise pricing changes, and no provider deprecating the model your prompts depend on.

My First User

I showed an early build to a friend who freelances as a web developer. They regularly pitch website modernization to local educational departments and businesses.

They entered their service pitch, plugged in a regional college's website, and generated an email. Their reaction:

"Usually AI writes stuff like 'I hope this email finds you in high spirits.' This actually noticed the recent curriculum update on their homepage and connected it to my web services. It reads like an email I spent 10 minutes writing myself."

They also told me what they want next: a batch upload feature, where they drop in a list of 50 links at night and wake up to 50 drafts.

That is now at the top of my roadmap.

Try It Yourself

bash
ollama pull qwen2.5:3b
git clone https://github.com/DeependraGurjar/coldcraft-studio.git
cd coldcraft-studio
pip install -r requirements.txt
python -m streamlit run app.py

Feedback, issues, and pull requests are welcome. If you have ideas for making local outreach tools better, I would love to hear them in the comments.

BioScavenger: A Quest Game That Talks for 30 Seconds, Then Sends Me Outside Without My Phone

Deependra Gurjar — Sun, 11 Oct 2026 16:07:01 +0000

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

What I Built

There’s an irony in most “go outside” apps: they keep you staring at your phone the whole time. I wanted to build the opposite.

BioScavenger is a small scavenger-hunt game you mostly play with your ears. You pick where you are (a park, your campus) and how much time you have, then tap Embark on Quest. A local AI writes a few simple tasks for you: touch the rough bark of a tree, close your eyes and pick out two different bird songs, find a leaf bigger than your hand. A calm voice reads them out like a short guided walk, starting with a deep breath.

Then you put your phone in your pocket and go.

When you’re back, you upload a photo of what you found, a flower or a leaf, and claim 50 Nature Credits per task. There are balloons. I’m not sorry.

I made it for anyone who needs a gentle excuse to step away from the screen, and who doesn’t want their location or photos sent to some cloud service in the process.

Code

GitHub: https://github.com/DeependraGurjar/Bioscavenger

The README walks through setting up Ollama, Piper, and Streamlit, and lists a few ideas for contributors.

How I Built It

Everything runs on my own laptop, on the CPU, no GPU.

The quests: Qwen2.5-3B through Ollama
The app asks a small local model for a quest and requests strict JSON back, with the narration and the tasks separated. That way the app can turn the answer into proper task cards instead of dumping a wall of text on the screen.

The voice: Piper TTS
The narration goes into Piper using the en_US-lessac-medium voice, which writes an audio file the app plays in the browser.

My first version was embarrassing. It was a 5-second robotic line that felt nothing like a guided walk. I rewrote the prompt to ask for a calm, mindfulness-style script of about 70–90 words, opening with a deep breath, and now the briefing runs 25–35 seconds. That one change made the whole app feel different.

The interface: Streamlit
Streamlit handles the dropdowns, the audio player, the photo uploader, and the credit counter. Finished tasks collapse so the page stays tidy, and claiming one sets off the balloons.

Keeping the repo small
Piper plus the voice model make my project folder about 98 MB, which is far too heavy for GitHub. I used .gitignore so only the code is pushed, and the README tells people how to download Piper and the voice themselves.

The Honest Part
The photo check is on the honor system. The app takes your upload and awards credits, but it doesn’t actually look at what’s in the picture. A local vision model could fix that, and it’s at the top of my contribution wishlist.
A 3B model is still a small model. It writes nice quests, but sometimes it suggests something that doesn’t fit the place. I treat its output as a prompt for a walk, not as fact.
“Offline” has fine print. Once the models and voice are downloaded, nothing needs the internet.

Why Open Innovation Matters Here
It all stays on my machine. The quest writing and the voice happen locally. No accounts, no API key, no tracking.

It costs nothing per quest. I can generate a new one every day and the only thing I spend is battery.

Anyone can change it. Because each piece is open and swappable, someone can try a different model, add a voice in their own language, or redesign the interface without asking permission.

And local AI isn’t magic:
First-time setup takes longer than calling a hosted API.
On a CPU, generation can feel slow. I added progress messages and a timeout so the app doesn’t look frozen.
Ollama gave me WinError 10061 connection errors, and I lost a while to that before I learned to check whether it was already running in the background.
The Walk
When and where: [Date, general place, how long]
The quest I got: [The actual tasks]
What worked: [A moment where the audio really made you notice something]
What didn’t: [A task that didn’t fit the place, or something off with the voice]
One surprise: [Something small you noticed outside]
Screen time: [What you really did with your phone]

Want to Help?
The repo is tagged for Hacktoberfest, and I’d love contributions. A few ideas:

Take photos straight from a phone camera
Narration in more languages
Real photo checking with a local vision model
Saving credits between sessions
Printable quest cards
Setup guides for Linux and macOS

DEV Community: Deependra Gurjar

SlothDesk: A Lazy, Offline AI That Cleans Your Downloads Folder

What I Built

What it does:

Code

DeependraGurjar / SlothDesk

Private, local-first AI file organizer. Reads PDF and text snippets offline via Ollama to automatically rename and sort your Downloads folder without cloud APIs.

🦥 SlothDesk Desktop Cleaner

Features

Categories

How I Built It

Why Does Open Innovation Matter?

Team

ColdCraft Studio: Personalized Cold Emails, Written Locally for $0

What I Built

Here is what happens when you click Generate:

Who is this for?

What happens

Key features

DeependraGurjar / coldcraft-studio

A local-first, zero-cost B2B cold outreach engine powered by Ollama and open-weight models.

ColdCraft Studio

Table of Contents

Overview

How I Built It

Why Open Innovation Matters

My First User

Try It Yourself

BioScavenger: A Quest Game That Talks for 30 Seconds, Then Sends Me Outside Without My Phone

What I Built

Code

How I Built It