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

What I Built

My Downloads folder is a pile of IMG_20260720_214610.jpg, Screenshot 2025-09-04 at 23.11.13.png and scan0003.pdf. I never sort it, because renaming and filing files by hand is boring and slow.

DownloadsSloth does the boring part. You point it at a folder, and Gemma 3 looks at each file and suggests a clearer name and a folder (Photos, Screenshots, Documents and so on). You see the whole plan, tick what you agree with, edit any name, and press one button. It copies the approved files into a new folder. Nothing is moved, renamed, overwritten or deleted, and there is an Undo button.

It runs entirely on my laptop. My files never leave the computer.

Demo

Screenshots from a real run on a copy of my Downloads folder:

Code

Setup: install Ollama, run ollama pull gemma3:4b, then npm install and npm start, and open http://127.0.0.1:3000.

The Lazy Payoff

What I measured, on a test folder of 44 files (images, PDFs, documents and spreadsheets copied from my own Downloads):

• Gemma produced suggestions for all 44 files in 141 seconds, unattended.
• After that, applying the approved changes took under a second for a batch.

What I did not measure: how long the same folder takes to sort by hand. I didn't time a manual run, so I won't claim a number. The saving is the typing and the opening of every file, and I can't say how big it is for you.

How I Built It

• Model: Gemma 3 4B through Ollama. For images it sees a small preview; for PDFs, documents and spreadsheets it uses only the file name, type and date, because the tool does not read their contents.
• Server: Node.js and Express, bound to 127.0.0.1 only. A background job processes files one at a time, with a progress bar and a Cancel button.
• Structured output: Gemma returns JSON (new name, folder, reason, confidence), validated and retried once.
• Built in stages: health check, read-only scan, suggestions, then apply and undo, testing each stage on a copy of my files before the next.

Safety Is Enforced in Code, Not Just the Prompt

A tool that touches your files has to be boring and careful, so these rules are in code:

• The scan is read-only, only works inside my home folder, and refuses system folders, ~/Library and hidden folders. I tested /, /etc, ~/.ssh, a relative path, and a ../../.. escape, and all were refused.
• The server re-scans the folder itself and never trusts file names sent from the browser. A path-traversal attempt (../../etc/passwd) is skipped as "not found".
• Apply only copies into a new output folder, never overwrites (a repeated name becomes -2), checks the size of every copy, and refuses an output folder that is the source or inside it.
• Undo removes only the copies the tool made, and only if their size still matches its log.
• The model never picks the extension or the final folder. Code keeps the original extension, cleans the name, and forces the folder to a fixed list.
• Low-confidence rows are unchecked by default.

What Went Wrong, and How I Fixed It

• Everything came back "High" confidence, even PDFs the tool had judged only by name. I capped confidence in code: non-images are never above "medium".
• It renamed files that were already clearly named. Now only meaningless names (IMG_..., Screenshot ..., scan0003, WhatsApp Image ...) are renamed. Clear names are kept and marked "kept".
• Wrong dates: a file from the evening of Oct 18 was labelled Oct 17, because the date came from UTC. It now uses the date in the file name, or the local date.
• Screenshots were called "portrait". They now go to a Screenshots folder.
• doc.pdf became 2025-12-07-pdf.pdf. When the tool can't read a file's content, it now keeps the name and sends the file to Review instead of inventing one.

Why Does Open Innovation Matter?

1. Privacy. These are the files people don't want to upload: bills, screenshots, résumés and photos. A hosted model would receive every one of them.
2. Cost. There is no per-file fee, so you can run it on a folder with thousands of files.
3. Control. I could wrap the model in checks and swap it later. Several of the fixes above exist because the 4B model ignored rules written in the prompt, so I enforced them in code.

What It Does Not Do

• The suggestions are guesses. A human reviews every change before anything is copied. One image was labelled "gangsters" with high confidence from the clothing in a still, which is not something a photo can prove.
• Documents are judged by name only. The tool does not read the inside of PDFs, documents or spreadsheets.
• It describes a picture, but doesn't know who is in it. I told it not to name people or characters.
• Image analysis is macOS only, because the preview uses the built-in sips tool.
• Small rough edges: an apply that copies zero files still creates an empty output folder, and characters like % in a name are turned into -.
• I did not time a manual run, I have no first user yet, and I have not tested it on Windows or Linux.

AI Disclosure

I used Antigravity to help write the code, and a Claude assistant helped me plan the build, test it and write this post. The app itself runs on Gemma locally.

Prize Categories

Best Use of Gemma: Gemma 3 4B suggests names and folders from image previews and file names, entirely on the user's own computer.