LazyDesk AI — Your Digital Sloth Assistant 🦥 (Stay Lazy. Get Organized.)

wpnews.pro

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

#

LazyDesk AI 🦥

Stay lazy. Get organized.

#

What I Built

Most people's computers have at least one terrifying directory—usually Downloads or Desktop—filled with chaotic files like:

- `Invoice_2026_Q1_draft (1).pdf`
- `Screenshot 2026-10-25 at 14.02.11.png`

Untitled_notes_final_v2.docx
expenses_backup (copy).xlsx Sorting through these files manually is tedious, repetitive, and mentally draining. You have to open every PDF just to figure out what company it's from, manually create folders, think of consistent filenames, and spot identical duplicates taking up space.

Inspired by the philosophy of the sloth—"do less manual work and let automation handle the boring chores"—I built LazyDesk AI.

LazyDesk AI is a privacy-first, desktop-class digital productivity assistant that:

Scans cluttered directories safely without touching system folders. 2. Finds exact duplicate files using fast size-filtered SHA-256 hashing and calculates reclaimable storage. 3. Categorizes files & suggests standardized filenames usingGoogle Gemma 2 (gemma2:2b) open weights. 4. Extracts executive document summaries from PDFs, Word docs, and Markdown with 5 key takeaways, deadlines, and action items. 5. Stages everything in a unified Action Plan with interactive approvals before touching a single byte on disk. 6. Journals all operations in SQLite with1-Click Undo to reverse file movements instantly if you change your mind.

Who gets to be a little lazier because of it?

Anyone drowning in a messy Downloads, Documents, or Desktop folder—students, freelancers, developers, researchers, and remote workers who want a clean, organized digital workspace without spending Sunday afternoon dragging icons into folders.

#

Demo

  🌐 Live Hosted Website (Render)

Try the interactive cloud demonstration directly in your browser:

👉 **[https://lazydesk-ai.onrender.com](https://lazydesk-ai.onrender.com)**

🚀 **[Deploy your own instance on Render with 1 Click](https://render.com/deploy?repo=https://github.com/akashgt/lazydesk-ai)**

Note on Hosted Demo: The live cloud deployment on Render runs in secure sandboxed mode (HOSTED_DEMO=true) with synthetic sample files (invoices, meeting notes, duplicated images). This lets you test the 1-click Lazy Mode, duplicate detection, and smart categorization safely in your browser without exposing your local filesystem.

⚡ Quick Run Locally (Full Desktop & File Access)

To run LazyDesk AI on your machine against your own messy folders with 100% offline privacy:

LazyDesk AI will launch the unified FastAPI server and automatically open your browser to http://127.0.0.1:8000/.

#

Code

  LazyDesk AI — Your Digital Sloth Assistant 🦥. Privacy-first local file organizer, duplicate finder, and summarizer powered by Google Gemma 2 and deployable on Render.

Stay lazy. Get organized.

LazyDesk AI is a privacy-first, desktop-class digital productivity assistant inspired by the philosophy of the sloth: do less manual work and let local open-weight automation handle the boring chores.

LazyDesk scans messy folders, categorizes files, suggests standardized filenames, detects exact duplicates, and generates document summaries with key takeaways—powered by Google's Gemma 2 open-weight model through Ollama, running 100% locally with zero cloud API keys, zero subscriptions, and complete data sovereignty.

LazyDesk AI also includes a production-ready Render Blueprint & multi-stage Docker deployment with sandboxed demo mode for hosted cloud demonstrations.

🏆 Partner Integrations #

1. Google Gemma — Best Use of Gemma

LazyDesk AI integrates Google's open-weight Gemma 2 (gemma2:2b) as its default local intelligence engine through Ollama:

Genuine Local Inference : Powers file classification into 7 smart categories, context-aware filename standardization, and executive document extraction with structured JSON output…

GitHub Repository :https://github.com/akashgt/lazydesk-ai #

License : MIT License (Application) • Google Gemma Terms of Use (Gemma 2 Model) #

Tech Stack : - Frontend : React 19, TypeScript, Vite, Tailwind CSS v4, Lucide Icons. #

Design System : Custom "Warm Ivory" theme (`#FAF8F5` ) with deep forest green accents (`#1B4332` ), soft sage cards, and expressive sloth mascots. #

Backend : FastAPI, Uvicorn, PyMuPDF (`fitz` ),`python-docx` , SQLite, Pydantic v2. #

Local AI Engine : Google Gemma 2 2B via Ollama HTTP Gateway. #

Cloud Infrastructure : Render Blueprint (render.yaml ), multi-stage Docker container,/api/health .

#

How I Built It

LazyDesk AI was built from the ground up around strict privacy, deterministic safety boundaries, and local open-weight AI.

Local AI Engine: Google Gemma 2 2B (gemma2:2b) Rather than relying on proprietary cloud APIs, LazyDesk uses Google's open-weight Gemma 2 2B (gemma2:2b) running locally through Ollama:

Fast, Bounded Inference : We tuned the Ollama gateway with bounded parameters (`num_ctx: 1536` ,`num_predict: 380` ,`temperature: 0.1` ,`timeout: 100.0s` ). This ensures deterministic JSON generation and snappy execution on standard consumer CPUs and laptops. #

Strict JSON Schema Prompts : Prompts include explicit JSON schemas and concrete output demonstrations. Responses are parsed and validated against strict Pydantic schemas. #

Flexible Model Adapter : The architecture preserves user sovereignty with a hot-swappable model adapter. Users can switch betweengemma2:2b ,qwen2.5:3b ,llama3.2:3b , or any installed Ollama model directly from the UI settings.

Prompt Hardening & Untrusted Data Boundary

A key design principle: document contents and filenames are untrusted data.

System prompts explicitly instruct the model that document text and filenames cannot override instructions, inject commands, or alter system behavior.
The model only proposes suggestions; pure deterministic Python code validates boundaries, checks path safety, and executes moves.

Path Traversal Defense & Collision Avoidance

Target paths are validated using:
Operations can never escape the selected directory.
Accidental file overwrites are impossible: a collision resolver checks if a target filename exists and automatically increments it (invoice_1.pdf ) non-destructively.

SQLite Safety Journal & 1-Click Rollback

Every file move is logged into an ACID-compliant SQLite journal (lazydesk_history.db) with source, destination, timestamp, and status. Clicking Undo Last Session walks backwards through the journal and restores all files to their exact original locations.

Render Cloud Blueprint & Multi-Stage Docker

To support cloud demonstration on Render:

`render.yaml` Blueprint : Configures a Web Service running on Render's Docker environment with automatic health checks (`/api/health` ) and dynamic port binding (`$PORT` ). #

Multi-Stage `Dockerfile` : Uses Node 20 to compile the React frontend into static assets, then packages the FastAPI backend into a slim Python 3.11 container. #

Hosted Demo Sandboxing : WhenHOSTED_DEMO=true is set, all file operations are strictly restricted to the syntheticsample_workspace/ directory.

#

Why Does Open Innovation Matter?

Open innovation is not just an alternative implementation detail for LazyDesk AI—it is the reason the application can ethically exist:

Privacy by Architecture : Desktop decluttering operates on your personal tax returns, medical bills, employment contracts, and private journals. Sending these confidential files across the internet to closed proprietary APIs creates massive privacy and compliance risks. Open-weight models like Gemma 2 keep private data 100% on your device. 2. Zero Per-Token Bills : Cleaning up 500 messy files or extracting summaries from twenty 30-page PDFs with closed APIs burns through tokens and demands recurring subscriptions. With local open innovation, intelligence is free and unlimited after down model weights. 3. True Offline Resilience : LazyDesk AI runs on airplanes, remote cabins, high-speed trains, or in air-gapped secure corporate facilities with zero internet access. 4. User Sovereignty & No Vendor Lock-in : If a closed API changes its pricing, deprecates an endpoint, or alters its content filters, apps break. With open weights, you own your tools and can inspect, modify, and run them indefinitely.

#

My Agent Session

This project was built pair-programming with the Antigravity AI Agent (Gemini 2.5 Flash / Pro):

The agent assisted across the full stack: architecting the FastAPI scanner, creating the warm ivory UI with vector sloth mascot illustrations, optimizing size-filtered SHA-256 duplicate detection, resolving Node 26 npm class-extension compatibility, testing real local Gemma 2 inference, and writing the Render deployment blueprint.
All 11 backend unit tests, frontend Vite builds, and live browser sessions were executed and verified directly within the Antigravity IDE workspace.

#

The Lazy Payoff

Decluttering a typical workspace folder with 50-100 messy downloads usually requires 20-40 minutes of clicking, creating folders, opening PDFs to see what they contain, and manually renaming files.

With LazyDesk AI:

Scan & Duplicate Detection : ~150 ms (local SHA-256 with byte-size pre-filtering). #

AI Categorization & Filename Suggestion : ~10-15 seconds for batch processing. #

Review & Execution : 1 click to apply the staged Action Plan. #

Undo : 1 click if you change your mind. #

Estimated Time Saved :~30 minutes per cleanup session , saving mental energy and eliminating digital clutter with zero effort.

#

Prize Categories

This project is entered into the following challenge tracks:

🦥 1. Hacktoberfest Open-Source AI Challenge — Week 3: Slow and Steady

“Stay lazy. Get organized.” Built to embody the sloth philosophy: automate tedious digital chores with local AI so humans can do less manual busywork.

💎 2. Best Use of Gemma (Google Gemma Partner Prize)

Integrates Google's open-weight Gemma 2 2B (gemma2:2b) as the default local AI engine via Ollama.
Uses real local Gemma 2 inference to categorize files into 7 intuitive groups, propose clean standardized filenames, and generate 5-bullet executive document summaries with deadlines and action items.
Features bounded contexts (num_ctx: 1536 ), deterministic JSON schemas, live model health diagnostics, and full compliance with the Google Gemma Terms of Use.

🚀 3. Best Use of Render (Render Partner Prize)

Includes a complete Infrastructure-as-Code Render Blueprint (render.yaml) and multi-stage container build ( Dockerfile ).
Implements automated zero-downtime health checking via /api/health and dynamic$PORT handling.
Features a secure Hosted Demo Mode (HOSTED_DEMO=true) that automatically sandboxes web visitors inside synthetic sample workspaces, proving that local-first desktop productivity apps can be showcased securely in the cloud on Render.

Stay lazy. Get organized with LazyDesk AI. 🦥

source & further reading

dev.to — original article Your AI-Built App Works. Here Is What Production Finds Anyway. I Built Termsight: An AI That Reads the Privacy Policy So You Don't Have To. Vector : local code summariser