DEV Community: Ayush Raj

LedgerGemma: Your Private AI Accountant — Powered by Open-Weight Gemma

Ayush Raj — Tue, 13 Oct 2026 06:19:00 +0000

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

🌿 LedgerGemma: Your Private AI Accountant

What if your AI accountant could organize your finances without sending your financial documents to a cloud AI service?

Freelancers, independent developers, and consultants deal with invoices, receipts, expense reports, and tax records almost every month. Organizing all that information manually is tedious, while cloud-based AI tools introduce potential privacy concerns and recurring costs.

I built LedgerGemma, a privacy-first expense and invoice management project designed to bring open-weight AI into everyday financial workflows.

My goal was simple: make financial document processing more accessible while keeping privacy and local control at the center of the experience.

💡 What I Built

LedgerGemma is a local-first expense and invoice intelligence application powered by open-weight Google Gemma models.

It combines a browser-based financial dashboard with a Python-based processing engine to help users organize financial information.

Here's what it aims to make easier:

🧾 Receipt and invoice processing: Extract structured financial information from document text.
🧠 AI-assisted categorization: Organize expenses into relevant categories using a Gemma-style prompt pipeline.
🔍 Expense auditing: Flag missing information and potential issues for further review.
📊 Financial exports: Export records into formats such as CSV and JSON for further analysis.
🔒 Privacy-first design: Keep browser records in local storage and support local processing workflows.
🌐 Progressive Web App: Access the browser-based interface through a responsive financial dashboard.

The project also includes a Python CLI engine for processing receipt text and a Node.js server that connects the browser interface with local processing functionality.

🎯 Who Is It For?

LedgerGemma is designed with freelancers, independent contractors, solo founders, and consultants in mind.

These users often need to organize business expenses without introducing unnecessary complexity into their financial workflows.

A privacy-focused tool can be particularly useful when documents contain confidential client or project information.

🚀 Demo

Source code: GitHub — LedgerGemma

Project README: Explore the architecture and setup instructions

To run the project locally:

git clone https://github.com/ayushraaj-01/ledgergemma.git
cd ledgergemma
npm install
npm start

Open the local address printed by the server in your browser.

You can also explore the Python processing engine:

python ledger_gemma_engine.py --demo

For a local receipt-text processing example:

python ledger_gemma_engine.py --scan path/to/receipt.txt

The CLI and browser application have different entry points, so you can explore the parts of the project that interest you most.

Note: The repository contains the source code and local setup instructions. A separately hosted, publicly accessible demo is not linked here.

🛠️ How I Built It

I wanted to explore how open-weight AI could become part of a practical application rather than remain a standalone model experiment.

LedgerGemma brings together several technologies.

1. Open-Weight AI with Gemma

The project uses Gemma-oriented conversation templates and structured-output instructions to guide the processing engine toward consistent financial records.

The prompt pipeline describes the information to extract, the expense categories to consider, and the structure expected from the model.

The repository is designed around compatibility with open-weight Gemma model workflows and local inference tools.

2. Python Processing Engine

The Python engine handles receipt-text processing and expense-related classification logic.

It provides a command-line interface, allowing users to experiment with processing financial documents independently of the browser UI.

3. Browser-Based Financial Dashboard

I used HTML, CSS, and JavaScript to build the browser interface.

The dashboard is designed to make financial records easier to inspect and manage, with local browser storage and data-export functionality.

4. Node.js Integration

A lightweight Express server serves the web application and exposes an endpoint for invoking the Python processing engine.

This provides a bridge between the browser experience and the Python-based workflow.

5. Privacy as a Design Principle

Rather than treating privacy as an optional feature, I designed the project around local data handling and local-first processing.

The intended workflow reduces dependence on third-party financial-data processing services.

The important distinction is that using open weights does not automatically guarantee offline operation or privacy. Those properties depend on how the model is executed, how the application is configured, and whether its components make external network requests.

🌍 Why Does Open Innovation Matter?

For me, open innovation means being able to understand, experiment with, and adapt the technology behind an application.

A closed, hosted AI API can be convenient, but it may introduce usage costs, infrastructure dependencies, and questions about where sensitive financial documents are processed.

An open-weight model creates different possibilities.

First, developers gain more control over deployment.

With a compatible local inference runtime, an application can be adapted to run on a user's own hardware instead of requiring every inference request to pass through a hosted model API.

Second, experimentation becomes more accessible.

Developers can inspect prompt templates, adapt extraction instructions, evaluate different model configurations, and improve the surrounding application without being tied to a single proprietary inference endpoint.

Third, privacy-conscious workflows become easier to explore.

Local processing can reduce the need to transmit financial documents to an external AI provider. This matters when invoices and receipts contain confidential business information.

Finally, open innovation makes the project easier to extend.

Other developers can contribute improved extraction logic, better document-processing workflows, additional export formats, and support for new model runtimes.

That is the part I find most exciting: the ability to build something practical while keeping the underlying technology open to experimentation and improvement.

🧪 What I Learned

Building LedgerGemma made me think beyond simply connecting an AI model to an interface.

Financial applications need more than plausible model responses. They need structured records, predictable data handling, validation, and ways for users to inspect the results.

I also explored how a browser dashboard, a Python processing engine, and an Express server can work together in a single project.

One of the biggest takeaways was that AI should assist financial workflows, not become the unquestioned authority on financial decisions.

Model-generated categories and potential tax classifications should be checked against reliable rules and reviewed by the user or a qualified professional before being used for actual tax filings.

🏆 Hacktoberfest Challenge

This project is my entry for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle.

I wanted to build something with a practical use case and explore how open-weight AI can support privacy-conscious software.

The project also explores the potential of Google's Gemma models and provides a starting point for further work on local AI-powered document processing.

If you're interested in open-weight AI, local-first applications, or financial automation, I'd love for you to explore the repository and share your feedback.

💻 Explore the Code

🌿 GitHub repository: https://github.com/ayushraaj-01/ledgergemma

Built with Python, JavaScript, HTML, CSS, Node.js, Express, and a Gemma-oriented AI processing pipeline.

What would you build with an open-weight AI model if you didn't have to start with a cloud API?

I'd love to hear your ideas. 🚀

StudyBuddy AI: A Personal Learning Companion Built with Open-Source AI

Ayush Raj — Mon, 12 Oct 2026 18:45:33 +0000

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

What I Built

Meet StudyBuddy AI — a personal AI learning companion designed to help my friends study smarter, manage their time, and make learning less stressful.

As college students, we often struggle with tight deadlines, difficult concepts, and balancing academics with coding practice. I wanted to build something that could make a real difference in a student's daily life.

StudyBuddy AI helps students:

📚 Understand complex topics through simple explanations.
📝 Generate quick summaries of study materials.
💻 Get help understanding programming concepts.
🎯 Break large study goals into manageable tasks.
🤖 Ask questions and receive personalized learning assistance.

The idea is simple: technology should help us learn better, not just work faster.

Demo

🚀 Live Demo: Coming Soon!

Code

💻 GitHub Repository: Coming Soon!

How I Built It

I designed the project around an AI-powered architecture with a focus on accessibility, flexibility, and an easy-to-use interface.

My planned tech stack includes:

Frontend: React.js and Tailwind CSS
Backend: Node.js and Express.js
AI: An open-weight language model
Inference: Ollama for running supported models locally
Version Control: Git and GitHub

The goal is to use an open-weight model for the core AI functionality, allowing users to experiment with different models and potentially run the application locally.

Why Does Open Innovation Matter?

Open innovation matters because AI should not be limited to people who can afford expensive APIs or powerful cloud infrastructure.

Building with open-weight models offers several advantages:

🔓 Freedom: Experiment with different models and configurations.
🔒 Privacy: Local inference can keep study materials and conversations on the user's device.
💰 Accessibility: Avoid per-request API charges when running models locally, although hardware and electricity still have costs.
🛠️ Customization: Adapt the system to different learning needs.

For me, the most exciting part is the possibility of creating useful AI tools without making users completely dependent on a closed platform.

My Agent Session

I'll share my development session here once it's available.

Prize Categories

Open Source AI

Final Thoughts

This project is about more than building another chatbot. It's about using open technology to solve a problem that students experience every day.

Sometimes, the best projects don't start with a groundbreaking idea. They start with a simple question:

"What can I build today that makes someone's life a little easier?"

That's the idea behind StudyBuddy AI. 🚀

VerdantEcho: A Zero-Signal Bioacoustics Trail Scout That Gets You Off the Screen

Ayush Raj — Sun, 11 Oct 2026 17:50:50 +0000

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

What I Built

VerdantEcho is an offline-first, zero-signal backcountry trail scout and bioacoustics neural classifier designed with a single goal: to make screen time the shortest part of your outdoor experience.

When hiking through deep mountain hollows, dense old-growth woods, or national parks, cellular reception completely drops. Cloud AI APIs fail with network timeouts. Worse, typical apps keep your eyes glued to glass while you walk past magnificent autumn canopies.

VerdantEcho changes this by running open-weight AI directly on-device:

🔊 On-Device Bioacoustics Scout: Uses the Web Audio API to render a live waterfall spectrogram and extracts spectral centroids and harmonic ratios to classify bird calls (Black-capped Chickadees, American Robins, Red-tailed Hawks, Song Sparrows) and autumn field crickets.
🍁 Autumn Foliage & Flora Vision: Analyzes canopy senescence, anthocyanin pigmentation ratios, and leaf morphology. Includes a dedicated hazard alert for Poison Ivy (Toxicodendron radicans) with high-contrast trail warnings.
❄️ Frost & Soil Almanac: Calculates first expected frost dates and tells you what to sow (hardneck garlic, winter spinach, winter rye) or harvest before the freeze based on offline date and coordinates.
🌲 "Touch Grass" Pocket Mode: Switches the device to an ultra-low-power AMOLED HUD. When an animal or plant is detected, synthesized speech announces the finding directly into your earphones, then instructs you to put the phone back in your pocket.
📍 Zero-Cloud Field Journal & GPX Exporter: Logs observations with offline GPS waypoints locally in IndexedDB and exports standard .gpx or .geojson files ready for Gaia GPS, AllTrails, or OpenStreetMap.
🤖 Gemma 2B Naturalist Companion: Formatted for Google's open-weight Gemma 2B model, delivering concise 2-sentence ecological insights paired with a firm reminder to put the screen away.

Demo

The app runs locally in the browser or offline on mobile devices via PWA service workers.

Key Screenshots & Live Features:

Real-time Bioacoustics Spectrogram: 2048-point FFT waterfall analyzer visualizes animal calls between 200 Hz and 12 kHz in real time.
Sunlight HUD & Pocket Mode: A high-contrast sunlight theme for bright daylight, plus an AMOLED Pocket Mode that automatically dims the screen and reads observations aloud.
Botanical Safety Detector: Instant warnings for trail hazards (Poison Ivy urushiol oil warnings) and autumn canopy color stages.

Code

ayushraaj-01 / verdantecho

Zero-signal offline trail scout & bioacoustics AI companion that gets you off the screen and into nature.

🌿 VerdantEcho: Zero-Signal Trail Scout & Bioacoustics AI

"Build something with open-source AI at its core that gets people off the screen and into the world. The best builds here should make the screen the shortest part of the experience."
— Hacktoberfest Open-Source AI Challenge: Week 1 (Touch Grass)

🍃 Overview

VerdantEcho is an offline-first, backcountry trail scout and bioacoustics neural classifier engineered for one primary purpose: to minimize your screen time and immerse you in the natural world.

When hiking through deep backcountry valleys, national parks, or remote woodlands, cell towers disappear. Cloud-based AI APIs (OpenAI, Claude, cloud computer vision) fail with timeout errors. Even worse, typical mobile apps encourage you to stare at a glowing piece of glass while walking through magnificent old-growth canopies.

VerdantEcho fundamentally re-engineers outdoor exploration using open-weight AI:

🔊 On-Device Bioacoustics Scout: Real-time Web Audio API FFT spectrogram + neural audio pattern extraction…

View on GitHub

GitHub Repository: https://github.com/ayushraaj-01/verdantecho

Includes complete source code:

index.html & styles.css: Vanilla HTML/CSS high-contrast forest design system.
app.js: Web Audio API FFT analyzer, canvas camera pipeline, and GPX exporter.
ai-worker.js: Web Worker for non-blocking bioacoustics inference.
gemma_field_scout.py: Python CLI for offline Gemma 2B turn-based field guidance.
sw.js: Service worker for 100% airplane-mode offline caching.
render.yaml & Dockerfile: 1-click cloud deployment configurations.

How I Built It

VerdantEcho was constructed using a lightweight, resilient, and offline-first architecture:

1. Web Worker Thread Isolation

Running audio spectrogram FFT calculations and image tensor analysis on the main UI thread causes dropped frames and unresponsiveness. Following proven patterns, we offloaded spectral math to an isolated Web Worker (ai-worker.js), ensuring silky smooth 60fps performance even on low-power mobile devices.

2. High-Contrast Sunlight & Pocket HUD

Standard web apps wash out under direct outdoor sunlight. VerdantEcho provides:

A high-contrast Sunlight HUD with dark emerald accents.
A Touch Grass Pocket Mode with automatic screen dimming, audio chimes, and Web Speech API synthesis so you never have to look down at the display.

3. Open-Weight Gemma 2B Naturalist Engine

We built a local Python CLI (gemma_field_scout.py) that demonstrates turn-based inference using Gemma's prompt format:


bash
python gemma_field_scout.py --species "Sugar Maple" --notes "High crimson saturation on ridge line"