DEV Community: Athul Jenson

Inbox Zen: Privacy-First File Sorting with Local Gemma

Athul Jenson — Fri, 23 Oct 2026 13:40:41 +0000

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

What I Built

"Inbox Zen" is a privacy-first AI agent that automatically categorizes and archives cluttered local files or emails overnight. It scans a designated "clutter" directory, reads the text, and intelligently sorts them into appropriate folders (like "Invoices", "Reading List", or "Personal Notes") using a local AI model. Finally, it generates a simple morning summary so you can wake up to a clean digital workspace without the dread of manually sorting files.

This directly targets the "Slow and Steady" theme: automating a tedious, repetitive daily chore.

Demo

Here is the live link to my daily "Inbox Zen" summary dashboard deployed on Render:
https://inbox-zen.onrender.com

(Note: If you visit, you'll see the latest summary of the files Gemma successfully sorted on my local machine!)

Code

You can find the full Python script and Flask server code in my GitHub repository here:
https://github.com/2AJ0/inbox-zen

How I Built It

I built this project completely manually focusing on a local-first approach to protect sensitive data:

AI Engine: I used Google's open-weight Gemma 2B model running completely offline via Ollama.
The Logic: A Python script uses the requests library to send file contents to the local Gemma instance, prompting it to strictly categorize the text. The script then uses shutil to move the files and generates a zen_summary.txt log.
Monitoring: I integrated the Sentry Python SDK to capture performance traces. I wrapped my main sorting function in a sentry_sdk.start_transaction() block and used .flush() to ensure trace data was sent before the script closed.
Deployment: I built a lightweight Flask web service to serve the text summary and deployed it directly from GitHub to Render.

Why Does Open Innovation Matter?

Open innovation is the absolute core of this project. Because I am dealing with personal files (like invoices or personal notes), sending this data to a closed, proprietary API in the cloud is a massive privacy risk. By utilizing an open-weight model like Gemma and running it locally with Ollama, my sensitive data never leaves my machine. It also means the tool is completely free to run with zero API costs, demonstrating how open-source AI democratizes powerful automation.

The Lazy Payoff

This agent saves me about 10-15 minutes of dragging and dropping files every single morning. More importantly, it removes the "digital tax" and mental friction of starting my day staring at a messy downloads folder. I just check my dashboard, see what was filed where, and start working.

Prize Categories

I am entering this submission for the following partner categories:

Best Use of Gemma (Primary focus: 100% local, privacy-first inference using Gemma 2B)
Render (Deployed the Flask summary dashboard)
Sentry (Integrated the Python SDK for performance tracing)

FetchUpdate – AI Pet Sitter Assistant

Athul Jenson — Fri, 16 Oct 2026 08:07:10 +0000

FetchUpdate – The AI Assistant for Local Pet Sitters 🐾

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

What I Built

FetchUpdate is an AI tool designed for independent dog walkers and local pet sitters.

Pet sitters often spend 1–2 hours every evening manually texting individual updates to anxious owners. FetchUpdate automates this chore: the sitter takes a daily photo of the pet, selects their profile, and the tool uses a vision model paired with stored pet care details to instantly draft a warm, highly personalized SMS update ready to copy and send.

Demo

👉 Live App: FetchUpdate on Streamlit Cloud

📁 GitHub Repository: github.com/2AJ0/fetchupdate

How I Built It

Database (MongoDB Atlas): Stores structured pet profiles, including names, breeds, special instructions (e.g., "Loves walks on paved trails"), owner details, and tone preferences.
Multimodal AI Pipeline (OpenRouter / Gemini / Open Models): Accepts the uploaded image along with the retrieved MongoDB profile and prompts an open-weight vision model to describe the activity and generate a tailored SMS update.
Frontend & Deployment (Streamlit): Built using Python and Streamlit, deployed seamlessly on Streamlit Community Cloud.

Why Does Open Innovation Matter?

For local side hustles and micro-businesses, open innovation makes software viable in two key ways:

Privacy First: Sitter notes contain sensitive client information (home locations, daily schedules, dog routines). Running open-weight vision models ensures client data stays secure without being harvested by closed corporate APIs.
Near-Zero Cost Structure: Closed API pricing quickly kills margins for small solo sitters charging $15–$30 per walk. Leveraging open models keeps operational overhead low enough to run a profitable local SaaS product.

My First User

I tested FetchUpdate with a local test profile for Merku, a Golden Retriever who loves long walks on paved trails.

Uploading a photo of Merku generated the following update in seconds:

"Hey Athul! Merku is sitting patiently, looking like the goodest boy. He's probably dreaming of soft treats and a nice long walk on the paved trails! 🐶"

The prompt correctly synthesized both the visual context from the photo and the background care instructions from MongoDB.

Prize Categories

Best Use of MongoDB Atlas: Used MongoDB Atlas to store long-term pet care instructions, owner metadata, and personality traits.
Best Use of Open Models / Gemma: Used open-weight vision models to generate contextual text updates from images.

Local Micro-Garden & Frost Date Advisor | Touch Grass Challenge

Athul Jenson — Thu, 08 Oct 2026 16:15:05 +0000

🌿 Overview: Bringing AI Outdoors ("Touch Grass")

Developers spend countless hours glued to screens and debugging code. For Week 1 of the Hacktoberfest Open-Source AI Challenge under the "Touch Grass" theme, I built the Local Micro-Garden & Frost Date Advisor Agent.

This interactive CLI agent evaluates regional frost limits, identifies cold-hardy crops, and creates an actionable outdoor task list based on how many hours you want to spend outside getting your hands in the soil.

🛠️ Tech Stack & Local Setup

The agent runs 100% locally on a Mac with zero external API dependencies or cloud key limits:

Framework: Mastra (@mastra/core) for agent orchestration and tool execution.
LLM Engine: Local llama3.1 model running via Ollama with native tool-calling support.
Type Safety & Validation: Zod for strict tool input/output schema enforcement.
Runtime: Node.js + TypeScript (tsx).

💡 Architecture & Tools

The agent coordinates two Mastra tools to build the garden plan:

getFrostDataTool: Analyzes historical frost dates and recommends cold-weather crops (such as Spinach, Garlic, Kale, and Winter Cabbage).
generatePlanTool: Converts your target outdoor time into a step-by-step checklist covering soil prep, mulching, and frost sheet installation.

🤖 Local Agent Definition (`src/agent.ts`)

import { Agent } from '@mastra/core/agent';
import { createOpenAI } from '@ai-sdk/openai';
import { getFrostDataTool, generatePlanTool } from './tools.js';

// Connect to local Ollama instance running llama3.1
const ollama = createOpenAI({
  baseURL: '[http://127.0.0.1:11434/v1](http://127.0.0.1:11434/v1)',
  apiKey: 'ollama',
});

export const gardenAgent = new Agent({
  id: 'micro-garden-agent',
  name: 'Micro-Garden Frost Advisor',
  instructions: `
    You are an AI assistant designed for the "Touch Grass" challenge.
    Your goal is to evaluate local frost dates and produce an outdoor weekend gardening plan.
    Encourage hands-on outdoor physical activity and practical soil maintenance.
  `,
  model: ollama('llama3.1'),
  tools: {
    getFrostData: getFrostDataTool,
    generatePlan: generatePlanTool,
  },
});

📸 Demo & Terminal Output
Running the CLI with Bengaluru and 5 outdoor hours:

🌱 --- Micro-Garden & Frost Date Advisor Agent --- 🌱

Enter your City or ZIP code: Bengaluru
How many hours do you want to spend outside gardening this weekend? 5

Consulting outdoor garden advisor...

--- YOUR WEEKEND OUTDOOR PLAN ---
Here is your customized outdoor action plan for Bengaluru:

Soil Prep: Apply a thick mulch layer to protect cold-hardy crops.
Planting Checklist: Plant Spinach, Garlic, and Winter Cabbage.
Frost Covers: Install simple protection sheets over sensitive beds.

Enjoy your 5 hours outdoors in the garden! 🌻
🔗 Repository & Quickstart

GitHub Repository: github.com/2AJ0/touch-grass-micro-garden-agent

How to Run Locally

# 1. Pull the tool-calling model in Ollama
ollama pull llama3.1

# 2. Clone & Install
git clone [https://github.com/2AJ0/touch-grass-micro-garden-agent.git](https://github.com/2AJ0/touch-grass-micro-garden-agent.git)
cd touch-grass-micro-garden-agent
npm install

# 3. Launch CLI Agent
npx tsx index.ts

What winter crops are you planning for your garden this season? Let me know in the comments! 🍃

DEV Community: Athul Jenson

Inbox Zen: Privacy-First File Sorting with Local Gemma

What I Built

Demo

Code

How I Built It

Why Does Open Innovation Matter?

The Lazy Payoff

Prize Categories

FetchUpdate – AI Pet Sitter Assistant

FetchUpdate – The AI Assistant for Local Pet Sitters 🐾

What I Built

Demo

How I Built It

Why Does Open Innovation Matter?

My First User

Prize Categories

Local Micro-Garden & Frost Date Advisor | Touch Grass Challenge

🌿 Overview: Bringing AI Outdoors ("Touch Grass")

🛠️ Tech Stack & Local Setup

💡 Architecture & Tools

🤖 Local Agent Definition (src/agent.ts)

GitHub Repository: github.com/2AJ0/touch-grass-micro-garden-agent

🤖 Local Agent Definition (`src/agent.ts`)